SIAnalytics/BMTT2022_SIA_track

BMTT2022

Release Note

CBAM and mixedstyle will be merged soon. Everything else has been uploaded.

bugfix : merge_json.py : 22-06-07

Install

Follow the official repo to install bytetrack.

Data Prepare

We used the MOTSynth official data extraction pipelines.

datasets
   |——————mot (MOT17)
   |        └——————train
   |        └——————test
   └——————motsynth
   |         └——————MOT17-02-DPM
   |         └——————MOT17-04-DPM
   |         └——————...
   |         └——————annotations
   |         └——————comb_annotations
   |         └——————frames
   └——————data_path

Training

To reproduce the performance, you need 8 GPUs with no less than 40G memory.

• Stage1. Training warm_up model with below script, or download warm-up model (58.1 HOTA), and save it in

python3 tools/train.py -f exps/example/mot/yolox_x_source_only.py -d 8 -b 48 --fp16 -o

• Make pseudo label, run below code

python3 tools/track.py -f exps/example/mot/yolox_x_mix_det.py -c weight/warm-up.pth.tar -b 1 -d 1 --fp16 --fuse
python3 tools/interpolation.py
python3 make_PU.py
python3 ./tools/convert_mot17_to_coco_pu.py
python3 merge_json.py

• Stage2. Cross-domain Mixed Sampling with mosaic augmentation

python3 tools/train.py -f exps/example/mot/yolox_x_mixed.py -d 8 -b 48 --fp16 -o -c weight/warm-up.pth.tar

• Make pseudo label, run below code

python3 tools/track.py -f exps/example/mot/yolox_x_ft.py -c weight/mixed.pth.tar -b 1 -d 1 --fp16 --fuse
python3 tools/interpolation.py
python3 make_PU.py
python3 ./tools/convert_mot17_to_coco_pu.py # We removed values with confidence less than 0.7 (L 108 in ./tools/convert_mot17_to_coco_pu.py) because predictions with low confidence can act as label noise.

• Stage3. Make multiple fine-tune model and model soup # when fine-tuned, the EMA is not used.

(Note that when performing fine-tune in Step 3, the augmentation combination should be different in L49-57 of ./yolox/data/datasets/mot.py)

python3 wa.py # you have to adjust it manually. (Until the CVPR22 conference, the completed code will be uploaded.)

Test

python3 tools/track.py -f exps/example/mot/yolox_x_source_only.py -c weight/warm-up_67.5.pth.tar -b 1 -d 1 --fp16 --fuse
python3 tools/interpolation.py # HOTA 58.1

python3 tools/track.py -f exps/example/mot/yolox_x_source_only.py -c weight/stage2_69.6.pth.tar -b 1 -d 1 --fp16 --fuse
python3 tools/interpolation.py # HOTA 59.xx

python3 tools/track.py -f exps/example/mot/yolox_x_source_only.py -c weight/stage3_75.7.pth.tar -b 1 -d 1 --fp16 --fuse
python3 tools/interpolation.py # HOTA 62.xx

python3 tools/track.py -f exps/example/mot/yolox_x_source_only.py -c weight/stage3_77.9.pth.tar -b 1 -d 1 --fp16 --fuse
python3 tools/interpolation.py # HOTA 63.xx

Sample