Code and Dataset for paper: Toward Pedestrian Head Tracking: A Benchmark Dataset and an Information Fusion Network
Cchead is an open dataset that is contributed by researchers and engineers from various colleges and companies. We appreciate all the contributors who implement their methods or add new features, as well as users who give valuable feedbacks. We wish that the dataset and benchmark could serve the growing research community by providing a flexible toolkit to reimplement existing methods and develop their own new detectors.
We present a Chinese Large-scale Cross-scene Pedestrian Head Tracking dataset (Cchead) and a Multi-Source Information Fusion Network (MIFN). Our dataset has features that are of considerable interest, including 10 diverse scenes of 50,528 frames with over 2,366,249 heads and 2,358 tracks annotated manually.
Dataset Download: will be open soon. We have submitted our Cchead dataset to Journal, and Cchead will be free downloaded after review. We will also release our trained Fairmot-mifn weights.
Video Demo: https://drive.google.com/drive/folders/1BLmzCRx3MbOzVUITw0-RCpRqTHJ2JXYQ?usp=sharing
We have uploaded both raw videos and filefolder of images and labels. If you want to replicate our experiment, you should unzip the 4 zipfiles into <yourpathtopaddledet/dataset/mot/>.
This filefolder contains what we directly applied in our methods.
It form the dataset as:
--balanced_data
--images
--train
--90cross_25fps
--img
--0.jpg
--1.jpg
--gt
--gt.txt
--seqinfo.ini
......
--test
--90cross_25fps
--img
--0.jpg
--1.jpg
--gt
--gt.txt
--seqinfo.ini
......
--labels_with_ids
--train
--90T_25fps
--img
--0.txt
--1.txt
......
--test
--90cross_25fps
--img
--0.txt
--1.txt
......
where gt.txt is used for evaluation, it is formed as:
1,1,57,86,28,32,1,1,1 (frame_id, track_id, xy (left&up) wh,...)
2,1,55,87,28,32,1,1,1
3,1,60,85,28,32,1,1,1
4,1,63,85,29,31,1,1,1
5,1,64,85,29,31,1,1,1
6,1,66,83,29,31,1,1,1
...
1,2,65,146,50,56,1,1,1 (frame_id, track_id, xy (left&up) wh,...)
label files for training such as 0.txt is formed as:
0 186 0.23671875 0.25069444444444444 0.0067708333333333336 0.015277777777777777 (class_ind, track_id, xy (center) wh)
.......
xywhs in label files are all scaled to (0,1)
seqinfo.ini contains the following information:
[Sequence] name=90T_25fps imDir=img frameRate=25 seqLength=1110 imWidth=1920 imHeight=1080 imExt=.jpg
If you are using paddledetection, you may modify codes of loading data to train network on this dataset.
- The code is tested on Ubuntu 20.04.2, python 3.9, cuda 11.7，Paddle 2.4.2, PaddleDet 2.6.
Please refer to [Installation]((https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/docs/tutorials/INSTALL_cn.md) for installation instructions of Paddle and PaddleDet. It is recommanded to install PaddleDet with pip3.
We modified some codes in PaddleDet, please clone them and replace the original code in PaddleDet.
python -m paddle.distributed.launch --log_dir=./fairmot_dla34_ourdata_30e_1088x608_mifn/ --gpus 0,1,2,3 tools/train.py -c configs/mot/fairmot/fairmot_ourdata_dla34_30e_1088x608_mifn.yml```CUDA_VISIBLE_DEVICES=1 python tools/eval_mot.py -c configs/mot/fairmot/fairmot_ourdata_dla34_30e_1088x608_mifn.yml -o weights=<path to weights>This work was sponsored by grants from National Science Foundation of China (Grant No. 52172308), Department of Transport of Hubei Province (Contract No. 2022- 11-4-4), National Natural Science Foundation of China under Grant No. 62192751 and the 111 International Collaboration Program of China under Grant No. BP2018006.