Multi-Granularity Tracking with Modularized Components for Unsupervised Vehicles Anomaly Detection (CVPRW 2020)
This repository contains source codes of team113 for NVIDIA AICity Challenge 2020 Track 4, and the technical details please refer to the paper "Multi-Granularity Tracking with Modularized Components for Unsupervised Vehicles Anomaly Detection"
Our method obtains the F1-score metric at 0.9855 and the RMSE metric at 4.8737, which ranked first in the Track4 test set of the NVIDIA AI CITY 2020 CHALLENGE.
Figure 1. The illustration of multi-granularity tracking with modularized components framework. This framework involves fusion from
box-level tracking branch and pixel-level tracking branch. The backtracking optimization is performed to further improve the predictions.
-
Paddle1.7-gpu-post97
-
cuda9
-
cudnn7.5
The training annotations for detection models are in here. Extract one frame every 2 seconds, then choose half of them to annotate. The corresponding is (annotation_number - 1)*2*30 = frame_number
cd det_code/PaddleDetection
sh train.sh
cd det_code/PaddleDetection
sh infer.sh
python bg_code/ex_bg_mog.py
python mask_code/mask_frame_diff.py
python mask_code/mask_track.py
python mask_code/mask_fuse.py
python box_track/tube_construction.py
python box_track/box_level_tracking.py
python pixel_track/coarse_ddet/pixel-level_tracking.py
python pixel_track/post_process/similar.py
python pixel_track/post_process/time_back.py
python pixel_track/post_process/id.py
python fusion_code/fusion_backtracking.py
contact us ([email protected] or [email protected])