1. Folder "frames". It contains 11,816 video frames captured by 6 cameras. For the naming rule of each frame, for example, in "c1s1_000151.jpg", "c1" refers to camera 1; "s1" refers to segment 1 in camera 1 (generated automatically during data collection); "000151" refers to the 151th frame in this segment. In addition, the train/test split is FIXED and provided by the files "frame_test.mat" and "frame_train.mat". 10-fold cross validation is NOT required.
2. Folder "annotations". It contains 11,816 MAT files in correspondence to each frame in folder "frames". All annotated boxes are pedestrians. Each MAT file records the bounding box position within the frame and its ID. The coordinates of each box are formatted in [x, y, w, h]. The ID of each box takes the value of [1, 932] as well as -2. "-2" means that we do not know for sure the ID of the person, and not used in the testing of person re-id, but is used in train/test of pedestrian detection (potentially used in the training of person re-identification).
3. Folder "query_box". It contains the query boxes of the PRW dataset. All togther there are 2057 queries. For naming rule, for example, in "479_c1s3_016471.jpg", "479" refers to the ID of the query, and "c1s3_016471" refers to the video frame where the query is cropped. Note that 1) the query IDs are not included in the training set, 2) the query images are not normalized (we typically use 128*64 for BoW extraction, and 224*224 for CNN feature extraction), 3)all queries are hand-drawn boxes, 4) we select one query image for each testing ID under each camera, so the maximum number of queries per ID is 6. In addition, we provide the bounding box information of each query in "query_info.txt", so one can generate the queries from the video frames through function "generate_query.m".
In addition, we provide the train/test split of the PRW dataset. One do not have to perform 10-fold cross validation. In detail, "frame_test.mat" and "frame_train.mat" specify the train/test frames, and "ID_test.mat" and "ID_train.mat" specify the train/test IDs. Note that a small portion of IDs used in training may appear in the testing frames, but will not appear in the testing IDs.
1. Pedestrian detection. Models can be trained on the training set specified in "frame_train" and tested on "frame_test". In the paper, we report both Average Precision and Miss Rate for pedestrian detection.
2. Person Re-identification. When you use your own detector or those provided by us, you will obtain your own gallery with varying sizes according to the threshold of the detectors. Then, with the given query images, standard person re-identification can be performed. We use mAP and CMC curve (and rank-1 accuracy) to evaluate overall re-identification performance. Note that, when your detector has poor performance, re-id accuracy will be low as a result.
One final note: PRW has the same origin with Market-1501, but the IDs are not consistent due to annotation issues.
Also, as the case in Market-1501, we will set up an leader board for PRW, and you are welcomed to submit your latest results and code to me and get exposed.
Please let me know if you have any questions or comments: liangzheng06@gmail.com.
