The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Development Kit

Mark Everingham - John Winn


Contents

1 Challenge

The goal of this challenge is to recognize objects from a number of visual object classes in realistic scenes (i.e. not pre-segmented objects). There are twenty object classes: There are five main tasks: In addition, there are two ``taster'' tasks:

2 Data

The VOC2012 data is released in two phases: (i) training and validation data with annotation is released with this development kit; (ii) test data without annotation is released at a later date.

2.1 Classification/Detection Image Sets

For the classification and detection tasks there are four sets of images provided:
train:
Training data
val:
Validation data (suggested). The validation data may be used as additional training data (see below).
trainval:
The union of train and val.
test:
Test data. The test set is not provided in the development kit. It will be released in good time before the deadline for submission of results.
Table 1 summarizes the number of objects and images (containing at least one object of a given class) for each class and image set. The data has been split into 50% for training/validation and 50% for testing. The distributions of images and objects by class are approximately equal across the training/validation and test sets. Note that the 2012 data for the main classification/detection tasks is the same as the 2011 data - no additional images have been annotated. The assignment of images to training/test sets has not been changed. The dataset includes images from the 2008-2011 datasets, for which no test set annotation has been released.
Table 1: Statistics of the main image sets. Object statistics list only the `non-difficult' objects used in the evaluation.

train val trainval test
  img obj img obj img obj img obj

Aeroplane		 327 432 343 433 670 865 - -
Bicycle 268 353 284 358 552 711 - -
Bird 395 560 370 559 765 1119 - -
Boat 260 426 248 424 508 850 - -
Bottle 365 629 341 630 706 1259 - -
Bus 213 292 208 301 421 593 - -
Car 590 1013 571 1004 1161 2017 - -
Cat 539 605 541 612 1080 1217 - -
Chair 566 1178 553 1176 1119 2354 - -
Cow 151 290 152 298 303 588 - -
Diningtable 269 304 269 305 538 609 - -
Dog 632 756 654 759 1286 1515 - -
Horse 237 350 245 360 482 710 - -
Motorbike 265 357 261 356 526 713 - -
Person 1994 4194 2093 4372 4087 8566 - -
Pottedplant 269 484 258 489 527 973 - -
Sheep 171 400 154 413 325 813 - -
Sofa 257 281 250 285 507 566 - -
Train 273 313 271 315 544 628 - -
Tvmonitor 290 392 285 392 575 784 - -
Total 5717 13609 5823 13841 11540 27450 - -

2.2 Segmentation Image Sets

For the segmentation task, corresponding image sets are provided as in the classification/detection tasks. To increase the amount of data, the training and validation image sets include images from the 2007-2011 segmentation tasks. The test set contains only 2008-2011 images (i.e. those for which no annotation has been released), and is a subset of the test set for the main tasks for which pixel-wise segmentations have been prepared. Table 2 summarizes the number of objects and images (containing at least one object of a given class) for each class and image set, for the combined 2007-2012 data. In addition to the segmented images for training and validation, participants are free to use the un-segmented training/validation images supplied for the main classification/detection tasks, and any annotation provided for the main challenge e.g. bounding boxes.
Table 2: Statistics of the segmentation image sets.

train val trainval test
  img obj img obj img obj img obj

Aeroplane		 88 108 90 110 178 218 - -
Bicycle 65 94 79 103 144 197 - -
Bird 105 137 103 140 208 277 - -
Boat 78 124 72 108 150 232 - -
Bottle 87 195 96 162 183 357 - -
Bus 78 121 74 116 152 237 - -
Car 128 209 127 249 255 458 - -
Cat 131 154 119 132 250 286 - -
Chair 148 303 123 245 271 548 - -
Cow 64 152 71 132 135 284 - -
Diningtable 82 86 75 82 157 168 - -
Dog 121 149 128 150 249 299 - -
Horse 68 100 79 104 147 204 - -
Motorbike 81 101 76 103 157 204 - -
Person 442 868 445 865 887 1733 - -
Pottedplant 82 151 85 171 167 322 - -
Sheep 63 155 57 153 120 308 - -
Sofa 93 103 90 106 183 209 - -
Train 83 96 84 93 167 189 - -
Tvmonitor 84 101 74 98 158 199 - -
Total 1464 3507 1449 3422 2913 6929 - -

2.3 Action Classification Image Sets

For the action classification task, corresponding image sets are provided as in the classification/detection tasks. The same images are used for both the action classification task (where a person is indicated by a bounding box) and the boxless action classification taster (where a person is indicated by a point lying somewhere on the body). Each person has been annotated with the set of actions they are performing from the set {jumping, phoning, playing a musical instrument, reading, riding a bicycle or motorcycle, riding a horse, running, taking a photograph, using a computer, walking}. Note that actions are not mutually exclusive, for example a person may simultaneously be walking and phoning. Additionally, there are people labeled with the action `other', meaning none of the above actions are being performed. These examples make serve as distractors to the main ten action classes. The image sets are disjoint from those of the classification/detection tasks and person layout taster task. Note that they are not fully annotated - only `person' objects forming part of the training and test sets are annotated, and there may be unannotated people in the images. Table 3 summarizes the action statistics for each image set.
Table 3: Statistics of the action classification image sets.

train val trainval test
  img obj img obj img obj img obj

Jumping		 204 247 201 248 405 495 - -
Phoning 223 228 221 229 444 457 - -
Playinginstrument 230 309 229 310 459 619 - -
Reading 232 264 231 266 463 530 - -
Ridingbike 201 289 199 289 400 578 - -
Ridinghorse 205 266 206 268 411 534 - -
Running 156 280 154 281 310 561 - -
Takingphoto 209 227 205 229 414 456 - -
Usingcomputer 199 237 196 239 395 476 - -
Walking 191 298 195 299 386 597 - -
Other 402 521 397 522 799 1043 - -
Total 2296 3134 2292 3144 4588 6278 - -

2.4 Person Layout Taster Image Sets

For the person layout taster task, corresponding image sets are provided as in the classification/detection tasks. A person is indicated by a bounding box, and each person has been annotated with part layout (head, hands, feet). Note that the 2012 person layout dataset is the same as the corresponding 2011 dataset - no additional images have been annotated. The assignment of images to training/test sets has not been changed. The dataset includes images from the 2008-2011 datasets, for which no test set annotation has been released, and the test set is disjoint from that for the main classification/detection tasks. Table 4 summarizes the number of `person' objects annotated with layout for each image set. Note that the person layout images in VOC2012 are not fully annotated - only a subset of people have been annotated, and no other objects - for example an image containing two people and a car may only have one person annotated.
Table 4: Statistics of the person layout taster image sets. Object statistics list only the `person' objects for which layout information (parts) is present.

train val trainval test
  img obj img obj img obj img obj

Person		 315 425 294 425 609 850 - -


2.5 Ground Truth Annotation

Objects of the twenty classes listed above are annotated in the ground truth. For each object, the following annotation is present: In preparing the ground truth, annotators were given a detailed list of guidelines on how to complete the annotation. These are available on the main challenge web-site [1]. Note that for the action classification images, only people have been annotated, and only the bounding box and a reference point on the person is available. Note also that for these images the annotation is not necessarily complete i.e. there may be unannotated people.

2.6 Segmentation Ground Truth

Figure 1: Example of segmentation ground truth. a. Training image b. Class segmentation showing background, car, horse and person labels. The cream-colored `void' label is also used in border regions and to mask difficult objects. c. Object segmentation where individual object instances are separately labelled.
\includegraphics[height=5cm]{inpseg.eps} \includegraphics[height=5cm]{clsseg.eps} \includegraphics[height=5cm]{objseg.eps}
a. b. c.
For the segmentation image sets, each image has two corresponding types of ground truth segmentation provided: Figure 1 gives an example of these two types of segmentation for one of the training set images. The ground truth segmentations are provided to a high degree of accuracy, but are not pixel accurate, as this would have greatly extended the time required to gather these segmentations. Instead, they were labelled so that a bordering region with a width of five pixels may contain either object or background. Bordering regions are marked with a `void' label (index 255), indicating that the contained pixels can be any class including background. The void label is also used to mask out ambiguous, difficult or heavily occluded objects and also to label regions of the image containing objects too small to be marked, such as crowds of people. All void pixels are ignored when computing segmentation accuracies and should be treated as unlabelled pixels during training. In addition to the ground truth segmentations given, participants are free to use any of the ground truth annotation for the classification/detection tasks e.g. bounding boxes.

2.7 Person Layout Taster Ground Truth

For the person layout taster task, `person' objects are additionally annotated with three `parts': For each annotated person, the presence or absence of each part is listed, and for each part present, the bounding box is specified. The test images for the person layout taster are disjoint from the other image sets. There are no `difficult' objects. As noted above, the layout taster ground truth is incomplete - only people are annotated, and there may be unannotated people in an image.

2.8 Action Classification Ground Truth

For the action classification task, `person' objects are annotated with bounding box, a reference point lying somewhere on the body, and a set of flags, one per action class e.g. `phoning' or `walking'. For each action the flag indicates if the person is performing the corresponding action. Note that actions are not mutually exclusive, for example a person may simultaneously be walking and phoning. The `other' action is mutually exclusive to all other actions. The image sets are disjoint from the classification/detection and layout taster tasks. There are no `difficult' objects.

3 Classification Task

3.1 Task

For each of the twenty object classes predict the presence/absence of at least one object of that class in a test image. The output from your system should be a real-valued confidence of the object's presence so that a precision/recall curve can be drawn. Participants may choose to tackle all, or any subset of object classes, for example ``cars only'' or ``motorbikes and cars''.

3.2 Competitions

Two competitions are defined according to the choice of training data: (i) taken from the VOC trainval data provided, or (ii) from any source excluding the VOC test data provided:
No. Task Training data Test data
1 Classification trainval test
2 Classification any but VOC test test
In competition 1, any annotation provided in the VOC train and val sets may be used for training, for example bounding boxes or particular views e.g. `frontal' or `left'. Participants are not permitted to perform additional manual annotation of either training or test data. In competition 2, any source of training data may be used except the provided test images. Researchers who have pre-built systems trained on other data are particularly encouraged to participate. The test data includes images from ``flickr'' (www.flickr.com); this source of images may not be used for training. Participants who have acquired images from flickr for training must submit them to the organizers to check for overlap with the test set.

3.3 Submission of Results

A separate text file of results should be generated for each competition (1 or 2) and each class e.g. `car'. Each line should contain a single identifier and the confidence output by the classifier, separated by a space, for example: 
comp1_cls_test_car.txt:
    ...
    2009_000001 0.056313
    2009_000002 0.127031
    2009_000009 0.287153
    ...
Greater confidence values signify greater confidence that the image contains an object of the class of interest. The example classifier implementation (section 9.2.1) includes code for generating a results file in the required format.


3.4 Evaluation

The classification task will be judged by the precision/recall curve. The principal quantitative measure used will be the average precision (AP). Example code for computing the precision/recall and AP measure is provided in the development kit. See also section 3.4.1. Images which contain only objects marked as `difficult' (section 2.5) are currently ignored by the evaluation. The final evaluation may include separate results including such ``difficult'' images, depending on the submitted results. Participants are expected to submit a single set of results per method employed. Participants who have investigated several algorithms may submit one result per method. Changes in algorithm parameters do not constitute a different method - all parameter tuning must be conducted using the training and validation data alone.


3.4.1 Average Precision (AP)

The computation of the average precision (AP) measure was changed in 2010 to improve precision and ability to measure differences between methods with low AP. It is computed as follows:
  1. Compute a version of the measured precision/recall curve with precision monotonically decreasing, by setting the precision for recall $r$ to the maximum precision obtained for any recall $r' \geq r$.
  2. Compute the AP as the area under this curve by numerical integration. No approximation is involved since the curve is piecewise constant.
Note that prior to 2010 the AP is computed by sampling the monotonically decreasing curve at a fixed set of uniformly-spaced recall values $0,0.1,0.2,\ldots,1$. By contrast, VOC2010-2012 effectively samples the curve at all unique recall values.

4 Detection Task

4.1 Task

For each of the twenty classes predict the bounding boxes of each object of that class in a test image (if any). Each bounding box should be output with an associated real-valued confidence of the detection so that a precision/recall curve can be drawn. Participants may choose to tackle all, or any subset of object classes, for example ``cars only'' or ``motorbikes and cars''.

4.2 Competitions

Two competitions are defined according to the choice of training data: (i) taken from the VOC trainval data provided, or (ii) from any source excluding the VOC test data provided:
No. Task Training data Test data
3 Detection trainval test
4 Detection any but VOC test test
In competition 3, any annotation provided in the VOC train and val sets may be used for training, for example bounding boxes or particular views e.g. `frontal' or `left'. Participants are not permitted to perform additional manual annotation of either training or test data. In competition 4, any source of training data may be used except the provided test images. Researchers who have pre-built systems trained on other data are particularly encouraged to participate. The test data includes images from ``flickr'' (www.flickr.com); this source of images may not be used for training. Participants who have acquired images from flickr for training must submit them to the organizers to check for overlap with the test set.

4.3 Submission of Results

A separate text file of results should be generated for each competition (3 or 4) and each class e.g. `car'. Each line should be a detection output by the detector in the following format: 
    <image identifier> <confidence> <left> <top> <right> <bottom>
where (left,top)-(right,bottom) defines the bounding box of the detected object. The top-left pixel in the image has coordinates $(1,1)$. Greater confidence values signify greater confidence that the detection is correct. An example file excerpt is shown below. Note that for the image 2009_000032, multiple objects are detected: 
comp3_det_test_car.txt:
    ...
    2009_000026 0.949297 172.000000 233.000000 191.000000 248.000000
    2009_000032 0.013737 1.000000 147.000000 114.000000 242.000000
    2009_000032 0.013737 1.000000 134.000000 94.000000 168.000000
    2009_000035 0.063948 455.000000 229.000000 491.000000 243.000000
    ...
The example detector implementation (section 9.2.2) includes code for generating a results file in the required format.

4.4 Evaluation

The detection task will be judged by the precision/recall curve. The principal quantitative measure used will be the average precision (AP) (see section 3.4.1). Example code for computing the precision/recall and AP measure is provided in the development kit. Detections are considered true or false positives based on the area of overlap with ground truth bounding boxes. To be considered a correct detection, the area of overlap $a_o$ between the predicted bounding box $B_p$ and ground truth bounding box $B_{gt}$ must exceed $50\%$ by the formula:
\begin{displaymath} a_o = \frac{area(B_p \cap B_{gt})}{area(B_p \cup B_{gt})} \end{displaymath} (1)

Example code for computing this overlap measure is provided in the development kit. Multiple detections of the same object in an image are considered false detections e.g. 5 detections of a single object is counted as 1 correct detection and 4 false detections - it is the responsibility of the participant's system to filter multiple detections from its output. Objects marked as `difficult' (section 2.5) are currently ignored by the evaluation. The final evaluation may include separate results including such ``difficult'' images, depending on the submitted results. Participants are expected to submit a single set of results per method employed. Participants who have investigated several algorithms may submit one result per method. Changes in algorithm parameters do not constitute a different method - all parameter tuning must be conducted using the training and validation data alone.

5 Segmentation Task

5.1 Task

For each test image pixel, predict the class of the object containing that pixel or 'background' if the pixel does not belong to one of the twenty specified classes. The output from your system should be an indexed image with each pixel index indicating the number of the inferred class (1-20) or zero, indicating background.

5.2 Competitions

Two competitions are defined according to the choice of training data: (i) taken from the VOC trainval data provided, or (ii) from any source excluding the VOC test data provided:
No. Task Training data Test data
5 Segmentation trainval test
6 Segmentation any but VOC test test
For competition 5, any annotation provided in the VOC train and val sets may be used for training, for example segmentation, bounding boxes or particular views e.g. `frontal' or `left'. However, if training uses annotation of any images other than the segmented training images, this must be reported as part of the submission (see below) since this allows a considerably larger training set. Participants are not permitted to perform additional manual annotation of either training or test data. For competition 6, any source of training data may be used except the provided test images.

5.3 Submission of Results

Submission of results should be as collections of PNG format indexed image files, one per test image, with pixel indices from 0 to 20. The PNG color map should be the same as the color map used in the provided training and validation annotation (MATLAB users can use VOClabelcolormap - see section 10.5.3). The example segmenter implementation (section 9.2.3) includes code for generating results in the required format. Participants may choose to include segmentations for only a subset of the 20 classes in which case they will be evaluated on only the included classes. For competition 5, along with the submitted image files, participants must also state whether their method used segmentation training data only or both segmentation and bounding box training data. This information will be used when analysing and presenting the competition results.

5.4 Evaluation

Each segmentation competition will be judged by average segmentation accuracy across the twenty classes and the background class. The segmentation accuracy for a class will be assessed using the intersection/union metric, defined as the number of correctly labelled pixels of that class, divided by the number of pixels labelled with that class in either the ground truth labelling or the inferred labelling. Equivalently, the accuracy is given by the equation,
$\displaystyle \mathrm{segmentation~accuracy} = \frac{\mathrm{true~positives}}{\mathrm{true~positives} + \mathrm{false~positives} + \mathrm{false~negatives}}$      

Code is provided to compute segmentation accuracies for each class, and the overall average accuracy (see section 10.5.2). Participants are expected to submit a single set of results per method employed. Participants who have investigated several algorithms may submit one result per method. Changes in algorithm parameters do not constitute a different method - all parameter tuning must be conducted using the training and validation data alone.

6 Action Classification Task

6.1 Task

For each of the ten action classes predict if a specified person (indicated by their bounding box) in a test image is performing the corresponding action. The output from your system should be a real-valued confidence that the action is being performed so that a precision/recall curve can be drawn. Participants may choose to tackle all, or any subset of action classes, for example ``walking only'' or ``walking and running''. Note that instances of the `other' action class are included in the training/test sets as negative examples; a classifier targeting the `other' action class is not required.

6.2 Competitions

Two competitions are defined according to the choice of training data: (i) taken from the VOC trainval data provided, or (ii) from any source excluding the VOC test data provided:
No. Task Training data Test data
9 Action Classification trainval test
10 Action Classification any but VOC test test
In competition 9, any annotation provided in the VOC train and val sets may be used for training. Participants may use images and annotation for any of the competitions for training e.g. horse bounding boxes/segmentation to learn `ridinghorse'. Participants are not permitted to perform additional manual annotation of either training or test data. In competition 10, any source of training data may be used except the provided test images. Researchers who have pre-built systems trained on other data are particularly encouraged to participate. The test data includes images from ``flickr'' (www.flickr.com); this source of images may not be used for training. Participants who have acquired images from flickr for training must submit them to the organizers to check for overlap with the test set.

6.3 Submission of Results

A separate text file of results should be generated for each competition (9 or 10) and each action class e.g. `phoning'. Each line should contain a single image identifier, object index, and the confidence output by the classifier, separated by a space, for example: 
comp9_action_test_phoning.txt:
    ...
    2010_006107  1  0.241236
    2010_006107  2  0.758739
    2010_006108  1  0.125374
    ...
The image identifier and object index specify the `person' object to which the output corresponds; these are provided in the corresponding image sets. Greater confidence values signify greater confidence that the person is performing the action of interest. The example implementation (section 9.2.4) includes code for generating a results file in the required format.


6.4 Evaluation

The action classification task will be judged by the precision/recall curve. The principal quantitative measure used will be the average precision (AP) (see section 3.4.1). Example code for computing the precision/recall and AP measure is provided in the development kit. Participants are expected to submit a single set of results per method employed. Participants who have investigated several algorithms may submit one result per method. Changes in algorithm parameters do not constitute a different method - all parameter tuning must be conducted using the training and validation data alone.

7 Boxless Action Classification Taster

7.1 Task

The boxless action classification task is the same as the action classification task, except that the person to be classified in a test image is indicated only by a single point lying somewhere on their body, instead of a tight bounding box. The aim is to evaluate the efficacy of action classification methods when they are not provided with ``precise'' information about the extent of the person (in the form a tight bounding box), as might be the case where the input to the method is obtained from a generic human detector. At training time, participants may use any of the annotation provided in the VOC dataset e.g. both reference point and bounding box for a person. At test time, a method must only use the reference point provided to identify the person to be classified - the bounding box present in the test image annotation must not be used.

7.2 Competitions

Two competitions are defined according to the choice of training data: (i) taken from the VOC trainval data provided, or (ii) from any source excluding the VOC test data provided:
No. Task Training data Test data
11 Action Classification trainval test (point only)
12 Action Classification any but VOC test test (point only)
In competition 11, any annotation provided in the VOC train and val sets may be used for training. Participants may use images and annotation for any of the competitions for training e.g. horse bounding boxes/segmentation to learn `ridinghorse'. Participants are not permitted to perform additional manual annotation of either training or test data. In competition 12, any source of training data may be used except the provided test images. Researchers who have pre-built systems trained on other data are particularly encouraged to participate. The test data includes images from ``flickr'' (www.flickr.com); this source of images may not be used for training. Participants who have acquired images from flickr for training must submit them to the organizers to check for overlap with the test set. Note that in both competitions the only annotation in the test images that may be used is the reference point for a person indicating which person is to be classified. It is not allowed to use the bounding boxes in the test annotation.

7.3 Submission of Results and Evaluation

Results should submitted in the same format as for the action classification task (using bounding boxes), using competition numbers 11 or 12 to identify that only reference points have been used for the test data. The taster will be evaluated in the same way as the action classification task.

8 Person Layout Taster

8.1 Task

For each `person' object in a test image (their bounding box is provided) predict the presence/absence of parts (head/hands/feet), and the bounding boxes of those parts. The prediction for a person layout should be output with an associated real-valued confidence of the layout so that a precision/recall curve can be drawn. Only a single estimate of layout should be output for each person. The success of the layout prediction depends both on: (i) a correct prediction of parts present/absent (e.g. is the hand visible or occluded); (ii) a correct prediction of bounding boxes for the visible parts.

8.2 Competitions

Two competitions are defined according to the choice of training data: (i) taken from the VOC trainval data provided, or (ii) from any source excluding the VOC test data provided:
No. Task Training data Test data
7 Layout trainval test
8 Layout any but VOC test test
In competition 7, any annotation provided in the VOC train and val sets may be used for training, for example bounding boxes or particular views e.g. `frontal' or `left'. Participants are not permitted to perform additional manual annotation of either training or test data. In competition 8, any source of training data may be used except the provided test images. Researchers who have pre-built systems trained on other data are particularly encouraged to participate. The test data includes images from ``flickr'' (www.flickr.com); this source of images may not be used for training. Participants who have acquired images from flickr for training must submit them to the organizers to check for overlap with the test set.

8.3 Submission of Results

To support the hierarchical (person+parts) nature of this task, an XML format has been adopted for submission of results. A separate XML file of results should be generated for each competition (6 or 7). The overall format should follow: 
<results>
    <layout>
        ... layout estimate 1 ...
    </layout>
    <layout>
        ... layout estimate 2 ...
    </layout>
</results>
Each detection is represented by a <layout> element. The order of detections is not important. An example detection is shown here: 
    <layout>
        <image>2009_000183</image>
        <object>1</object>
        <confidence>-1189</confidence>
        <part>
            <class>head</class>
            <bndbox>
                <xmin>191</xmin>
                <ymin>25</ymin>
                <xmax>323</xmax>
                <ymax>209</ymax>
            </bndbox>
        </part>
        <part>
            <class>hand</class>
            <bndbox>
                <xmin>393</xmin>
                <ymin>206</ymin>
                <xmax>488</xmax>
                <ymax>300</ymax>
            </bndbox>
        </part>
        <part>
            <class>hand</class>
            <bndbox>
                <xmin>1</xmin>
                <ymin>148</ymin>
                <xmax>132</xmax>
                <ymax>329</ymax>
            </bndbox>
        </part>
    </layout>
The <image> element specifies the image identifier. The <object> specifies the index of the object to which the layout relates (the first object in the image has index 1) and should match that provided in the image set files (section 10.1.6). The <confidence> element specifies the confidence of the layout estimate, used to generate a precision/recall curve as in the detection task. Each <part> element specifies the detection of a particular part of the person i.e. head/hand/foot. If the part is predicted to be absent/invisible, the corresponding element should be omitted. For each part, the <class> element specifies the type of part: head, hand or foot. The <bndbox> element specifies the predicted bounding box for that part; bounding boxes are specified in image co-ordinates and need not be contained in the provided person bounding box. To ease creation of the required XML results file for MATLAB users, a function is included in the development kit to convert MATLAB structures to XML. See the VOCwritexml function (section 10.7.1). The example person layout implementation (section 9.2.6) includes code for generating a results file in the required format.

8.4 Evaluation

The person layout task will principally be judged by how well each part individually can be predicted. For each of the part types (head/hands/feet) a precision/recall curve will be computed, using the confidence supplied with the person layout to determine the ranking. A prediction of a part is considered true or false according to the overlap test, as used in the detection challenge, i.e. for a true prediction the bounding box of the part overlaps the ground truth by at least 50%. For each part type, the principal quantitative measure used will be the average precision (AP) (see section 3.4.1). Example code for computing the precision/recall curves and AP measure is provided in the development kit.

8.4.0.1 Invitation to propose evaluation schemes.

We invite participants to propose additional evaluation schemes for the person layout task. In particular, we are interested in schemes which (i) evaluate accuracy of complete layout predictions; (ii) incorporate the notion of ranking of results by confidence. If you have a successful layout prediction method and insight please propose promising evaluation techniques in the form of (i) motivation and explanation; (ii) MATLAB implementation compatible with the VOC results format.

9 Development Kit

The development kit is packaged in a single gzipped tar file containing MATLAB code and (this) documentation. The images, annotation, and lists specifying training/validation sets for the challenge are provided in a separate archive which can be obtained via the VOC web pages [1].

9.1 Installation and Configuration

The simplest installation is achieved by placing the development kit and challenge databases in a single location. After untarring the development kit, download the challenge image database and untar into the same directory, resulting in the following directory structure: 
    VOCdevkit/                           % development kit
    VOCdevkit/VOCcode/                   % VOC utility code
    VOCdevkit/results/VOC2012/           % your results on VOC2012
    VOCdevkit/local/                     % example code temp dirs
    VOCdevkit/VOC2012/ImageSets          % image sets
    VOCdevkit/VOC2012/Annotations        % annotation files
    VOCdevkit/VOC2012/JPEGImages         % images
    VOCdevkit/VOC2012/SegmentationObject % segmentations by object
    VOCdevkit/VOC2012/SegmentationClass  % segmentations by class
If you set the current directory in MATLAB to the VOCdevkit directory you should be able to run the example functions: If desired, you can store the code, images/annotation, and results in separate directories, for example you might want to store the image data in a common group location. To specify the locations of the image/annotation, results, and working directories, edit the VOCinit.m file, e.g. 
% change this path to point to your copy of the PASCAL VOC data
VOCopts.datadir='/homes/group/VOCdata/';

% change this path to a writable directory for your results
VOCopts.resdir='/homes/me/VOCresults/';

% change this path to a writable local directory for the example code
VOCopts.localdir='/tmp/';
Note that in developing your own code you need to include the VOCdevkit/VOCcode directory in your MATLAB path, e.g. 
>> addpath /homes/me/code/VOCdevkit/VOCcode

9.2 Example Code

Example implementations are provided for all tasks. The aim of these (minimal) implementations is solely to demonstrate use of the code in the development kit.


9.2.1 Example Classifier Implementation

The file example_classifier.m contains a complete implementation of the classification task. For each VOC object class a simple classifier is trained on the train set; the classifier is then applied to the val set and the output saved to a results file in the format required by the challenge; a precision/recall curve is plotted and the `average precision' (AP) measure displayed.


9.2.2 Example Detector Implementation

The file example_detector.m contains a complete implementation of the detection task. For each VOC object class a simple (and not very successful!) detector is trained on the train set; the detector is then applied to the val set and the output saved to a results file in the format required by the challenge; a precision/recall curve is plotted and the `average precision' (AP) measure displayed.


9.2.3 Example Segmenter Implementation

An example segmenter is provided which converts detection results into segmentation results, using create_segmentations_from_detections (described below). For example: 
>> example_detector;
>> example_segmenter;
This runs the example detector, converts the detections into segmentations and displays a table of per-class segmentation accuracies, along with an overall average accuracy.


9.2.4 Example Action Implementation

The file example_action.m contains a complete implementation of the action classification task. For each VOC action class a simple classifier is trained on the train set; the classifier is then applied to all specified `person' objects in the val set and the output saved to a results file in the format required by the challenge; a precision/recall curve is plotted and the `average precision' (AP) measure displayed.


9.2.5 Example Boxless Action Implementation

The file example_action_nobb.m contains a complete implementation of the boxless action classification task, using only the reference point annotation of people in the test images (no bounding boxes). For each VOC action class a simple classifier is trained on the train set; the classifier is then applied to all specified `person' objects in the val set and the output saved to a results file in the format required by the challenge; a precision/recall curve is plotted and the `average precision' (AP) measure displayed.


9.2.6 Example Layout Implementation

The file example_layout.m contains a complete implementation of the person layout task. A simple (and not very successful!) layout predictor is trained on the train set; the layout predictor is then applied to the val set and the output saved to a results file in the format required by the challenge; a precision/recall curve is plotted and the `average precision' (AP) measure displayed.

9.3 Non-MATLAB Users

For non-MATLAB users, the file formats used for the VOC2012 data should be straightforward to use in other environments. Image sets (see below) are vanilla text files. Annotation files are XML format and should be readable by any standard XML parser. Images are stored in JPEG format, and segmentation ground truth in PNG format.

10 Using the Development Kit

The development kit provides functions for loading annotation data. Example code for computing precision/recall curves and segmentation accuracy, and for viewing annotation is also provided.

10.1 Image Sets

10.1.1 Classification/Detection Task Image Sets

The VOC2012/ImageSets/Main/ directory contains text files specifying lists of images for the main classification/detection tasks. The files train.txt, val.txt, trainval.txt and test.txt list the image identifiers for the corresponding image sets (training, validation, training+validation and testing). Each line of the file contains a single image identifier. The following MATLAB code reads the image list into a cell array of strings: 
imgset='train';
ids=textread(sprintf(VOCopts.imgsetpath,imgset),'%s');
For a given image identifier ids{i}, the corresponding image and annotation file paths can be produced thus: 
imgpath=sprintf(VOCopts.imgpath,ids{i});
annopath=sprintf(VOCopts.annopath,ids{i});
Note that the image sets used are the same for all classes. For each competition, participants are expected to provide output for all images in the test set.

10.1.2 Classification Task Image Sets

To simplify matters for participants tackling only the classification task, class-specific image sets with per-image ground truth are also provided. The file VOC2012/ImageSets/Main/<class>_<imgset>.txt contains image identifiers and ground truth for a particular class and image set, for example the file car_train.txt applies to the `car' class and train image set. Each line of the file contains a single image identifier and ground truth label, separated by a space, for example: 
    ...
    2009_000040 -1
    2009_000042 -1
    2009_000052  1
    ...
The following MATLAB code reads the image list into a cell array of strings and the ground truth label into a corresponding vector: 
imgset='train';
cls='car';
[ids,gt]=textread(sprintf(VOCopts.clsimgsetpath, ...
                    cls,imgset),'%s %d');
There are three ground truth labels:
-1:
Negative: The image contains no objects of the class of interest. A classifier should give a `negative' output.
1:
Positive: The image contains at least one object of the class of interest. A classifier should give a `positive' output.
0:
``Difficult'': The image contains only objects of the class of interest marked as `difficult'.
The ``difficult'' label indicates that all objects of the class of interest have been annotated as ``difficult'', for example an object which is clearly visible but difficult to recognize without substantial use of context. Currently the evaluation ignores such images, contributing nothing to the precision/recall curve or AP measure. The final evaluation may include separate results including such ``difficult'' images, depending on the submitted results. Participants are free to omit these images from training or include as either positive or negative examples.

10.1.3 Segmentation Image Sets

The VOC2012/ImageSets/Segmentation/ directory contains text files specifying lists of images for the segmentation task. The files train.txt, val.txt, trainval.txt and test.txt list the image identifiers for the corresponding image sets (training, validation, training+validation and testing). Each line of the file contains a single image identifier. The following MATLAB code reads the image list into a cell array of strings: 
imgset='train';
ids=textread(sprintf(VOCopts.seg.imgsetpath,imgset),'%s');
For a given image identifier ids{i}, file paths for the corresponding image, annotation, segmentation by object instance and segmentation by class can be produced thus: 
imgpath=sprintf(VOCopts.imgpath,ids{i});
annopath=sprintf(VOCopts.annopath,ids{i});
clssegpath=sprintf(VOCopts.seg.clsimgpath,ids{i});
objsegpath=sprintf(VOCopts.seg.instimgpath,ids{i});
Participants are expected to provide output for all images in the test set.

10.1.4 Action Classification Image Sets

The VOC2012/ImageSets/Action/ directory contains text files specifying lists of images and `person' objects for the action classification tasks. The files train.txt, val.txt, trainval.txt and test.txt list the image identifiers for the corresponding image sets (training, validation, training+validation and testing). Each line of the file contains a single image identifier. The following MATLAB code reads the image list into a cell array of strings: 
imgset='train';
ids=textread(sprintf(VOCopts.action.imgsetpath,imgset),'%s');
For a given image identifier ids{i}, the corresponding image and annotation file paths can be produced thus: 
imgpath=sprintf(VOCopts.imgpath,ids{i});
annopath=sprintf(VOCopts.annopath,ids{i});
Note that the image sets used are the same for all action classes. For each competition, participants are expected to provide output for all `person' objects in each image of the test set.

10.1.5 Action Class Image Sets

To simplify matters for participants tackling the action classification task, action class-specific image sets with per-object ground truth are also provided. The file VOC2012/ImageSets/Action/<class>_<imgset>.txt contains image identifiers, object indices and ground truth for a particular action class and image set, for example the file phoning_train.txt applies to the `phoning' action class and train image set. Each line of the file contains a single image identifier, single object index, and ground truth label, separated by a space, for example: 
    ...
    2010_006215  1  1
    2010_006217  1 -1
    2010_006217  2 -1
    ...
The following MATLAB code reads the image identifiers into a cell array of strings, the object indices into a vector, and the ground truth label into a corresponding vector: 
imgset='train'; cls='phoning';
[imgids,objids,gt]=textread(sprintf(VOCopts.action.clsimgsetpath,
                                   cls,imgset),'%s %d %d');
The annotation for the object (bounding box, reference point and actions) can then be obtained using the image identifier and object index: 
    rec=PASreadrecord(sprintf(VOCopts.annopath,imgids{i}));
    obj=rec.objects(objids{i});
There are two ground truth labels:
-1:
Negative: The person is not performing the action of interest. A classifier should give a `negative' output.
1:
Positive: The person is performing the action of interest. A classifier should give a `positive' output.


10.1.6 Person Layout Taster Image Sets

The VOC2012/ImageSets/Layout/ directory contains text files specifying lists of image for the person layout taster task. The files train.txt, val.txt, trainval.txt and test.txt list the image identifiers for the corresponding image sets (training, validation, training+validation and testing). Each line of the file contains a single image identifier, and a single object index. Together these specify a `person' object for which layout is provided or to be estimated, for example: 
    ...
    2009_000595  1
    2009_000595  2
    2009_000606  1
    ...
The following MATLAB code reads the image list into a cell array of strings and the object indices into a corresponding vector: 
imgset='train';
[imgids,objids]=textread(sprintf(VOCopts.layout.imgsetpath, ...
                            VOCopts.trainset),'%s %d');
The annotation for the object (bounding box only in the test data) can then be obtained using the image identifier and object index: 
    rec=PASreadrecord(sprintf(VOCopts.annopath,imgids{i}));
    obj=rec.objects(objids{i});

10.2 Development Kit Functions

10.2.1 VOCinit

The VOCinit script initializes a single structure VOCopts which contains options for the PASCAL functions including directories containing the VOC data and options for the evaluation functions (not to be modified). The field classes lists the object classes for the challenge in a cell array: 
VOCopts.classes={'aeroplane','bicycle','bird','boat',...
                 'bottle','bus','car','cat',...
                 'chair','cow','diningtable','dog',...
                 'horse','motorbike','person','pottedplant',...
                 'sheep','sofa','train','tvmonitor'};
The field actions lists the action classes for the action classification task in a cell array: 
VOCopts.actions={'other','jumping','phoning','playinginstrument',...
                 'reading','ridingbike','ridinghorse','running',...
                 'takingphoto','usingcomputer','walking'};
The field trainset specifies the image set used by the example evaluation functions for training: 
VOCopts.trainset='train'; % use train for development
Note that participants are free to use both training and validation data in any manner they choose for the final challenge. The field testset specifies the image set used by the example evaluation functions for testing: 
VOCopts.testset='val'; % use validation data for development
Other fields provide, for convenience, paths for the image and annotation data and results files. The use of these paths is illustrated in the example implementations.

10.2.2 PASreadrecord(filename)

The PASreadrecord function reads the annotation data for a particular image from the annotation file specified by filename, for example: 
>> rec=PASreadrecord(sprintf(VOCopts.annopath,'2009_000067'))

rec =

       folder: 'VOC2009'
     filename: '2009_000067.jpg'
       source: [1x1 struct]
         size: [1x1 struct]
    segmented: 0
      imgname: 'VOC2009/JPEGImages/2009_000067.jpg'
      imgsize: [500 334 3]
     database: 'The VOC2009 Database'
      objects: [1x6 struct]
The imgname field specifies the path (relative to the main VOC data path) of the corresponding image. The imgsize field specifies the image dimensions as (width,height,depth). The database field specifies the data source (e.g. VOC2009 or VOC2012). The segmented field specifies if a segmentation is available for this image. The folder and filename fields provide an alternative specification of the image path, and size an alternative specification of the image size: 
>> rec.size

ans =

     width: 500
    height: 334
     depth: 3
The source field contains additional information about the source of the image e.g. web-site and owner. This information is obscured until completion of the challenge. Objects annotated in the image are stored in the struct array objects, for example: 
>> rec.objects(2)

ans =

        class: 'person'
         view: 'Right'
    truncated: 0
     occluded: 0
    difficult: 0
        label: 'PASpersonRight'
     orglabel: 'PASpersonRight'
         bbox: [225 140 270 308]
       bndbox: [1x1 struct]
      polygon: []
         mask: []
     hasparts: 1
         part: [1x4 struct]
The class field contains the object class. The view field contains the view: Frontal, Rear, Left (side view, facing left of image), Right (side view, facing right of image), or an empty string indicating another, or un-annotated view. The truncated field being set to 1 indicates that the object is ``truncated'' in the image. The definition of truncated is that the bounding box of the object specified does not correspond to the full extent of the object e.g. an image of a person from the waist up, or a view of a car extending outside the image. Participants are free to use or ignore this field as they see fit. The occluded field being set to 1 indicates that the object is significantly occluded by another object. Participants are free to use or ignore this field as they see fit. The difficult field being set to 1 indicates that the object has been annotated as ``difficult'', for example an object which is clearly visible but difficult to recognize without substantial use of context. Currently the evaluation ignores such objects, contributing nothing to the precision/recall curve. The final evaluation may include separate results including such ``difficult'' objects, depending on the submitted results. Participants may include or exclude these objects from training as they see fit. The bbox field specifies the bounding box of the object in the image, as [left,top,right,bottom]. The top-left pixel in the image has coordinates $(1,1)$. The bndbox field specifies the bounding box in an alternate form: 
>> rec.objects(2).bndbox

ans =

    xmin: 225
    ymin: 140
    xmax: 270
    ymax: 308
For backward compatibility, the label and orglabel fields specify the PASCAL label for the object, comprised of class, view and truncated/difficult flags. The polygon and mask fields specify polygon/per-object segmentations, and are not provided for the VOC2012 data. The hasparts field specifies if the object has sub-object ``parts'' annotated. For the VOC2012 data, such annotation is available for a subset of the `person' objects, used in the layout taster task. Object parts are stored in the struct array part, for example: 
>> rec.objects(2).part(1)

ans =

        class: 'head'
         view: ''
    truncated: 0
     occluded: 0
    difficult: 0
        label: 'PAShead'
     orglabel: 'PAShead'
         bbox: [234 138 257 164]
       bndbox: [1x1 struct]
      polygon: []
         mask: []
     hasparts: 0
         part: []
The format of object parts is identical to that for top-level objects. For the `person' parts in the VOC2012 data, parts are not annotated with view, or truncated/difficult flags. The bounding box of a part is specified in image coordinates in the same way as for top-level objects. Note that the object parts may legitimately extend outside the bounding box of the parent object. For `person' objects in the action classification image sets, objects are additionally annotated with the set of actions being performed and a reference point on the person's body. The hasactions field specifies if the object has actions annotated. Action flags are stored in the struct actions, for example: 
>> rec.objects(1).actions

ans =

                other: 0
              phoning: 1
    playinginstrument: 0
              reading: 0
           ridingbike: 0
          ridinghorse: 0
              running: 0
          takingphoto: 0
        usingcomputer: 0
              walking: 0
There is one flag for each of the ten action classes plus `other', with the flag set to true (1) if the person is performing the corresponding action. Note that actions except `other' are not mutually-exclusive. Each person in the action classification image sets is additionally annotated with a reference point which is used to indicate the person's approximate location for the boxless action classification task. The haspoint field specifies if the object has a reference point annotated, which is stored in the struct point, for example: 
>> rec.objects(1).point

ans =

    x: 186
    y: 210
The point is guaranteed to lie on the body and to be unoccluded by other objects. Typically the point is located around the middle of their chest, although this varies depending on pose and occlusions.

10.2.3 viewanno(imgset)

The viewanno function displays the annotation for images in the image set specified by imgset. Some examples: 
>> viewanno('Main/train');
>> viewanno('Main/car_val');
>> viewanno('Layout/train');
>> viewanno('Segmentation/val');
>> viewanno('Action/trainval');

10.3 Classification Functions

10.3.1 VOCevalcls(VOCopts,id,cls,draw)

The VOCevalcls function performs evaluation of the classification task, computing a precision/recall curve and the average precision (AP) measure. The arguments id and cls specify the results file to be loaded, for example: 
>> [rec,prec,ap]=VOCevalcls(VOCopts,'comp1','car',true);
See example_classifier for further examples. If the argument draw is true, the precision/recall curve is drawn in a figure window. The function returns vectors of recall and precision rates in rec and prec, and the average precision measure in ap.

10.4 Detection Functions

10.4.1 VOCevaldet(VOCopts,id,cls,draw)

The VOCevaldet function performs evaluation of the detection task, computing a precision/recall curve and the average precision (AP) measure. The arguments id and cls specify the results file to be loaded, for example: 
>> [rec,prec,ap]=VOCevaldet(VOCopts,'comp3','car',true);
See example_detector for further examples. If the argument draw is true, the precision/recall curve is drawn in a figure window. The function returns vectors of recall and precision rates in rec and prec, and the average precision measure in ap.

10.4.2 viewdet(id,cls,onlytp)

The viewdet function displays the detections stored in a results file for the detection task. The arguments id and cls specify the results file to be loaded, for example: 
>> viewdet('comp3','car',true)
If the onlytp argument is true, only the detections considered true positives by the VOC evaluation measure are displayed.

10.5 Segmentation Functions

10.5.1 create_segmentations_from_detections(id,confidence)

This function creates segmentation results from detection results. create_segmentations_from_detections(id) creates segmentations from the detection results with specified identifier e.g. comp3. This is achieved by rendering the bounding box for each detection in class order, so that later classes overwrite earlier classes (e.g. a person bounding box will overwrite an overlapping an aeroplane bounding box). All detections will be used, no matter what their confidence level. create_segmentations_from_detections(id,confidence) does the same, but only detections above the specified confidence will be used. See example_segmenter for an example.


10.5.2 VOCevalseg(VOCopts,id)

The VOCevalseg function performs evaluation of the segmentation task, computing a confusion matrix and segmentation accuracies for the segmentation task. It returns per-class percentage accuracies, the average overall percentage accuracy, and a confusion matrix, for example: 
>> [accuracies,avacc,conf,rawcounts] = VOCevalseg(VOCopts,'comp3')
Accuracies are defined by the intersection/union measure. The optional fourth output `rawcounts' returns an un-normalized confusion matrix containing raw pixel counts. See example_segmenter for another example. This function will also display a table of overall and per-class accuracies.


10.5.3 VOClabelcolormap(N)

The VOClabelcolormap function creates the color map which has been used for all provided indexed images. You should use this color map for writing your own indexed images, for consistency. The size of the color map is given by N, which should generally be set to 256 to include a color for the `void' label.

10.6 Action Functions

10.6.1 VOCevalaction(VOCopts,id,cls,draw)

The VOCevalaction function performs evaluation of the action classification task, computing a precision/recall curve and the average precision (AP) measure. The arguments id and cls specify the results file to be loaded, for example: 
>> [rec,prec,ap]=VOCevalcls(VOCopts,'comp9','phoning',true);
See example_action for further examples. If the argument draw is true, the precision/recall curve is drawn in a figure window. The function returns vectors of recall and precision rates in rec and prec, and the average precision measure in ap. Note that the same evaluation function applies to both the action classification task and boxless action classification task.

10.7 Layout Functions


10.7.1 VOCwritexml(rec,path)

The VOCwritexml function writes a MATLAB structure array to a corresponding XML file. It is provided to support the creation of XML results files for the person layout taster. An example of usage can be found in example_layout.

10.7.2 VOCevallayout_pr(VOCopts,id,draw)

The VOCevallayout_pr function performs evaluation of the person layout task, computing a precision/recall curve and the average precision (AP) measure for each part type (head/hands/feet). The arguments id and cls specify the results file to be loaded, for example: 
>> [rec,prec,ap]=VOCevallayout_pr(VOCopts,'comp7',true);
See example_layout for further examples. If the argument draw is true, the precision/recall curves are drawn in a figure window. The function returns vectors of recall and precision rates in reci and prec{i}, and the average precision measure in ap{i}, where the index i indexes the part type in VOCopts.parts.

Acknowledgements

We gratefully acknowledge the following, who spent many long hours providing annotation for the VOC2012 database: Yusuf Aytar, Lucia Ballerini, Hakan Bilen, Ken Chatfield, Mircea Cimpoi, Ali Eslami, Basura Fernando, Christoph Godau, Bertan Gunyel, Phoenix/Xuan Huang, Jyri Kivinen, Markus Mathias, Kristof Overdulve, Konstantinos Rematas, Johan Van Rompay, Gilad Sharir, Mathias Vercruysse, Vibhav Vineet, Ziming Zhang, Shuai Kyle Zheng. The preparation and running of this challenge is supported by the EU-funded PASCAL2 Network of Excellence on Pattern Analysis, Statistical Modelling and Computational Learning.

Bibliography

1
The PASCAL Visual Object Classes Challenge (VOC2012).
http://pascallin.ecs.soton.ac.uk/challenges/VOC/voc2012/index.html.

About this document ...

The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Development Kit

This document was generated using the LaTeX2HTML translator Version 2002 (1.62)

Copyright © 1993, 1994, 1995, 1996, Nikos Drakos, Computer Based Learning Unit, University of Leeds.
Copyright © 1997, 1998, 1999, Ross Moore, Mathematics Department, Macquarie University, Sydney.

The command line arguments were:
latex2html -split 0 -no_navigation -show_section_numbers devkit_doc.tex

The translation was initiated by me on 2012-05-21

me 2012-05-21