CTMC-v1-solution

This repostitory contains the 2021 CTMC-v1 (Cell Tracking With Mitosis Detection Dataset Challenge) solution - CTMC-v1 is part of the MOT Challenge.

More details and step by step expalantion can be found in this (Polish only) blog post.

We follow a tracking-by-detection approach

• detection is done using a trained YOLOv4 model
• we provide examples with three different tracking methods:
  • IoU Tracking
  • V-IoU Tracking
  • norfair.

Sample sequence frame

Instructions

Clone the repository

git clone https://github.com/deepdrivepl/CTMC-v1-solution.git
cd CTMC-v1-solution

Install the requirements

conda create -n CTMC python=3.7
conda activate CTMC
pip install -r requirements.txt

Run inference (weights)

darknet detector test cfgs/011-v4.data cfgs/011-v4.cfg backups/011-v4_best.weights -ext_output -dont_show -out results/011-best-val.json < filelists/011-val

Convert detections to MOT format

python MOT/yolo2MOT/yolo2MOT.py --input results/011-best-val.json --out_dir results/011-best-val/detection --json

Run IoU Tracking

git clone https://github.com/bochinski/iou-tracker.git
cd iou-tracker
while read p; do python demo.py --detection_path ../results/011-best-val/detection/$p.txt --output_path ../results/011-best-val/iou/$p.txt --format motchallenge; done < ../filelists/011-val-seq

Run V-IoU Tracking

To run V-IoU Tracking, you need to download training images and save them in the train directory.

while read p; do python demo.py --detection_path ../results/011-best-val/detection/$p.txt --output_path ../results/011-best-val/viou/$p.txt --format motchallenge --frames_path ../train/$p/img1/{:06d}.jpg --visual KCF; done < ../filelists/011-val-seq

Run tracking with norfair

Jupyter notebook

Visualize

python MOT/visualizeMOT.py --imgs train --mot_dir results/011-best-val/iou --out_dir results/011-best-val/iou --img_mot
python MOT/visualizeMOT.py --imgs train --mot_dir results/011-best-val/viou --out_dir results/011-best-val/viou --img_mot
python MOT/visualizeMOT.py --imgs train --mot_dir results/011-best-val/norfair --out_dir results/011-best-val/norfair --img_mot

Compute metrics

We use py-motmetrics.

git clone https://github.com/cheind/py-motmetrics.git && cd py-motmetrics
python -m motmetrics.apps.eval_motchallenge ../train ../results/011-best-val/iou
python -m motmetrics.apps.eval_motchallenge ../train ../results/011-best-val/viou
python -m motmetrics.apps.eval_motchallenge ../train ../results/011-best-val/norfair

Results

	IDF1	IDP	IDR	Rcll	Prcn	GT	MT	PT	ML	FP	FN	IDs	FM	MOTA	MOTP	IDt	IDa	IDm
OVERALL IoU	46.5%	49.6%	43.7%	64.2%	72.9%	374	137	178	59	54413	81299	1864	4629	39.5%	0.295	115	1747	55
OVERALL v-IoU	45.4%	48.4%	42.7%	64.2%	72.8%	374	137	178	59	54528	81297	1968	4975	39.4%	0.296	107	1867	52
OVERALL norfair	52.0%	54.1%	50.0%	65.3%	70.6%	374	134	184	56	61925	78829	1107	3280	37.6%	0.298	118	894	64

Output video (YOLOv4 + norfair)

U2O-S-run03.mp4

Acknowledgements

• https://github.com/bochinski/iou-tracker
• https://github.com/tryolabs/norfair
• https://github.com/cheind/py-motmetrics
• https://github.com/AlexeyAB/darknet