In this repo, we include the 1st Place submission to Aicity Challenge 2020 vehicle counting track(Baidu submission).
Zhongji Liu, Wei Zhang, Xu Gao, Hao Meng, Zhan Xue, Xiao Tan, Xiaoxing Zhu, Hongwu Zhang, Shilei Wen, and Errui Ding. Robust movement-specific vehicle counting at crowded intersections. In Proc. CVPR Workshops, Seattle, WA, USA, 2020. Contact: [email protected], [email protected]. Any questions or discussion are welcome!
AICITY2020 Challange Track1 Leaderboard
| TeamName | Score |
|---|---|
| Baidu(Ours) | 0.9389 |
| ENGIE | 0.9346 |
| CMU | 0.9292 |
| BUT | 0.8829 |
| KISTI | 0.8540 |
This repo is developed under the following configurations:
-
Hareware: Centos 7, 4 NVIDIA P4 GPUs
-
Software: Python=3.6.7 GCC 5.4.0 paddle=1.7.0 CUDA=9.0
We provide a docker image for official test for track-1 datasetB.
Download via https://bj.bcebos.com/v1/baixue/liuzhongji/aicity2020_task1_counting.tar
We also provide our label of datasetA for fineturning vehicle detector.
Download via Baidu Cloud: [https://pan.baidu.com/s/1o8QMCubHCO6dTTSELCcNuA passwd: r08r]
where:
train.tar -- selected frames from datasetA
1584811984_train_final.json -- label
faster_rcnnn_r50_track1.tar -- configurations for fineturning
####How to run:
- load the docker image and create a docker container(or prepare paddle environment follow this)
docker load < baidu_aicity2020_track1.tar
nvidia-docker run --name aicity2020_task1 --shm-size 16G -it aicity2020:track1
docker attach aicity2020_track1
the code is in the path: /home/task1_code, the cuda env file is in the path: /home/cuda
-
prepare the dataset( copy the whole dataset of "AIC20_track1" to
/home/task1_code*) -
prepare the run env:
cd /home/task1_code && source set_env.sh -
modify the path of list_video_id.txt in: (1) run_pipeline.sh ; (2) vehicle_counting/counting.py
-
run the whole pipline:
sh run_pipeline.sh
the steps and related folders included in the whole pipeline:
- **extract_frames.sh: **extract images from videos in dataset like "AIC20_track"
- **run_detection.sh: ** /home/task1_code/PaddleDetection is the folder where we run detection model, the origional /tools/infer.py, which you can find at paddle detection is modified to save detection results in txt format. The detection results will be saved in /home/task1_code/det_results/
- **run_tracking.sh: ** /home/task1_code/tracking is the folder where we run online tracking method, the tracking results will be saved to txt files in /home/task1_code/track_results/
- **run_counting.sh: **/home/task1_code/vehicle_counting is the folder wher we run counting method, the final counting results will be saved to txt files in /home/task1_code/vehicle_counting_results/
The related repos:
[paddle detection] https://github.com/PaddlePaddle/PaddleDetection/tree/release/0.2
[py-hausdorff] https://github.com/mavillan/py-hausdorff