Third IEEE International Workshop on Performance Evaluation of Tracking and Surveillance (PETS'2002)

PETS2002 Datasets

Dataset

The image sequences (and corresonding images) for PETS'2002 may be obtained from here.

Important instructions are given at the bottom of this page on processing the datasets - please read these carefully.

If you would like the datasets to be sent to you (CDROM format) please complete the form below.

PETS'2002 consists of datasets for two distinct applications:

- people tracking and counting

- hand posture classification

MOVING PEOPLE

The datasets consists of people moving in front of a shop window. The task is to determine:

- how many people are passing in front of the shop window
- how many people stop and look into the window
- how many people are looking into the window at each instant (frame) in time
- the trajectories of people passing in front of the store
- the time spent per frame (processing time): a histogram of the microseconds spent processing each frame.

It is recommended that these results (for test datasets 1, 2, and 3) are placed into graph(s)/table in the paper submission.

The task is made difficult by:

- the reflections from the ground and from the opposing window
- occlusions from the text on the window
- occluding groups of people walking together
- occlusions among the people standing in front of the store

The datasets are available under here

The entire people training set (gzipped tar file) can be downloaded here.

The entire people test set (gzipped tar file) can be downloaded here.

You can also download all files under one directory using wget i.e.
wget -m ftp://pets.rdg.ac.uk/PETS2002/PEOPLE/
will download all people datasets to the current directory. Please see http://www.gnu.org/software/wget/wget.html for more details.

Note that both individual Jpeg files and MPEG sequences are available for each dataset.

If you are having problems, please access via direct ftp as shown above.

There are two directories - one for training, and one for testing. Results should be presented only on the test data - see below for more detailed instructions.

PEOPLE - TRAINING: DATASET 1 (377 frames)

Frame 107 Frame 191
People tracking, dataset 1, frame 107

PEOPLE - TRAINING: DATASET 2 (1471 frames)

Frame 665 Frame 967

PEOPLE - TRAINING: DATASET 3 (1297 frames)

Frame 83 Frame 770

PEOPLE - TESTING: DATASET 1 (653 frames)

Frame 390 Frame 510

PEOPLE - TESTING: DATASET 2 (1752 frames)

Frame 1116 Frame 1240

PEOPLE - TESTING: DATASET 3 (1304 frames)

Frame 570 Frame 730

Notes on datasets:

the training directory contains a training set of frames in both MPEG movie format and as individual Jpeg images for two training sets.
the test directory contains a training set of frames in both MPEG movie format and as individual Jpeg images for four test sets.

The tracking/counting can be performed on the entire test sequence 1, 2, or 3, or a portion of it. The images may be converted to any other format as appropriate, e.g. subsampled, or converted to monochrome. All results reported in the paper should clearly indicate which part of the test data is used, ideally with reference to frame numbers where appropriate, e.g. PEOPLE TRACKING, TEST DATASET 1, FRAMES 1-100.

There is no requirement to report results on all the test sequences (1, 2, and 3), however you are encouraged to test your algorithms on as much of the test data as possible.

Please ensure when quoting training or test sequences in the paper, that they match the datasets on this website.

The tracking results must be submitted along with the paper, with the tracking results generated in XML format (details to be posted here soon.) This will be straightforward and should not add a significant overhead to your effort.

Acknowledgement
The sequences have been provided by the consortium of Project IST VISOR BASE (IST-1999-10808) - http://www.vtools.es/visorbase/index.html

HAND POSTURE RECOGNITION

The datasets consist of close-ups of hand postures in front of a uniform light, dark and complex backgrounds. The task is to perform hand posture classification. A protocol package is provided (see below) giving details about which data to use for training anf testing.

Note: currently there exist no reference databases and no standards for the evaluation and comparison of developed algorithms in hand posture recognition, and more generally in gesture recognition.

The datasets for PETS2002 are the Jochen Triesch Hand Posture Gallery available here.
The individual PGM files may be viewed here.

The protocol package including details on training and testing is provided here.

If you are having problems, please access via direct ftp as shown above.

Details on how to submit your classification results in XML format will be added here shortly.

VERY IMPORTANT INFORMATION FOR PAPER SUBMISSION

The results that you report in your paper submission (irrespective of either application - hand posture recognition or people tracking/counting):

should be performed using the test data, but the training data may optionally be used if the algorithms require it for learning etc.

The paper that you submit may be based on previously published tracking methods/algorithms (including papers submitted to the main ECCV conference). The importance is that your paper MUST report results using the datasets supplied.

If you have any queries please email pets2002@visualsurveillance.org.