tov

Scale Match for TinyPerson Detection (WACV2020)

Prerequisites

Install enviroment

As Install said.

prepare dataset

TinyPerson

to train baseline of TinyPerson, download the mini_annotation of all annotation is enough, which can be downloaded as tiny_set/mini_annotations.tar.gz in Baidu Yun(password:pmcq) / Google Driver.

mkdir data
ln -s ${Path_Of_TinyPerson} data/tiny_set
tar -zxvf data/tiny_set/mini_annotations.tar.gz && mv mini_annotations data/tiny_set/

dataset download link

Official Site: recomended, download may faster
Baidu Pan password: pmcq
Google Driver
For more details about TinyPerson dataset, please see Dataset.

COCO

ln -s ${Path_Of_COCO} data/coco

Experiment

TinyPerson

For running more experiment, to see bash script

# exp1.2: Faster-FPN, 2GPU
export GPU=2 && LR=0.01 && CUDA_VISIBLE_DEVICES=0,1 PORT=10000 tools/dist_train.sh configs2/TinyPerson/base/faster_rcnn_r50_fpn_1x_TinyPerson640.py $GPU \
  --work-dir ../TOV_mmdetection_cache/work_dir/TinyPerson/Base/faster_rcnn_r50_fpn_1x_TinyPerson640/old640x512_lr${LR}_1x_${GPU}g/ \
  --cfg-options optimizer.lr=${LR}

# Scale Match
# exp4.0 coco-sm-tinyperson: Faster-FPN, coco batch=8x2
export GPU=2 && export LR=0.01 && export BATCH=8 && export CONFIG="faster_rcnn_r50_fpn_1x_coco_sm_tinyperson"
export COCO_WORK_DIR="../TOV_mmdetection_cache/work_dir/TinyPerson/scale_match/${CONFIG}/lr${LR}_1x_${BATCH}b${GPU}g/"
CUDA_VISIBLE_DEVICES=0,1 PORT=10001 tools/dist_train.sh configs2/TinyPerson/scale_match/${CONFIG}.py $GPU \
  --work-dir ${COCO_WORK_DIR} --cfg-options optimizer.lr=${LR} data.samples_per_gpu=${BATCH}
# python exp/tools/extract_weight.py ${COCO_WORK_DIR}/latest.pth
export TCONFIG="faster_rcnn_r50_fpn_1x_TinyPerson640"
export GPU=2 && LR=0.01 && CUDA_VISIBLE_DEVICES=0,1 PORT=10000 tools/dist_train.sh configs2/TinyPerson/base/${TCONFIG}.py $GPU \
  --work-dir ../TOV_mmdetection_cache/work_dir/TinyPerson/scale_match/${TCONFIG}/cocosm_old640x512_lr${LR}_1x_${GPU}g/ \
  --cfg-options optimizer.lr=${LR} load_from=${COCO_WORK_DIR}/latest.pth

• GPU: 3080 x 2
• Adap RetainaNet-c means use clip grad while training.
• COCO val $mmap$ only use for debug, cause val also add to train while sm/msm coco to TinyPerson

the result is with max_det=200, it not right because there maybe exists 800+ objects in a single object, so we re-evaluate(not finished) with max_det=1000 and got following result

detector	type	$AP_{50}^{tiny}$	script	COCO200 val $mmap$	coco batch/lr
Faster-FPN	-	47.90 49.81	configs2/TinyPerson/base/Baseline_TinyPerson.sh:exp1.2	-	-
Faster-FPN	SM	50.06 50.85	ScaleMatch_TinyPerson.sh:exp4.0	18.9	8x2/0.01
Faster-FPN	SM	49.53 50.30	ScaleMatch_TinyPerson.sh:exp4.1	18.5	4x2/0.01
Faster-FPN	MSM	49.39 50.18	ScaleMatch_TinyPerson.sh:exp4.2	12.1	4x2/0.01
--	--	--
Adap RetainaNet-c	-	43.66 45.22	configs2/TinyPerson/base/Baseline_TinyPerson.sh:exp2.3	-	-
Adap RetainaNet-c	SM	50.07 51.78	ScaleMatch_TinyPerson.sh:exp5.1	19.6	4x2/0.01
Adap RetainaNet-c	MSM	48.39 50.00	ScaleMatch_TinyPerson.sh:exp5.2	12.9	4x2/0.01

for more experiment, to see TinyPerson experiment for detail of scale match, to see TinyPerson Scale Match

Citation

If you use the code and benchmark in your research, please cite:

@inproceedings{yu2020scale,
  title={Scale Match for Tiny Person Detection},
  author={Yu, Xuehui and Gong, Yuqi and Jiang, Nan and Ye, Qixiang and Han, Zhenjun},
  booktitle={The IEEE Winter Conference on Applications of Computer Vision},
  pages={1257--1265},
  year={2020}
}

And if the ECCVW challenge sumarry do some help for your research, please cite:

@article{yu20201st,
  title={The 1st Tiny Object Detection Challenge: Methods and Results},
  author={Yu, Xuehui and Han, Zhenjun and Gong, Yuqi and Jan, Nan and Zhao, Jian and Ye, Qixiang and Chen, Jie and Feng, Yuan and Zhang, Bin and Wang, Xiaodi and others},
  journal={arXiv preprint arXiv:2009.07506},
  year={2020}
}