Towards Resolving the Challenge of Long-tail Distribution in UAV Images for Object Detection

Introduction

This repo is the official implementation for WACV 2021 paper: Towards Resolving the Challenge of Long-tail Distribution in UAV Images for Object Detection. .

Requirements

1. Environment:

The implementation is based on mmdetection. So the requirements are exactly the same as mmdetection v2.3.0rc0+8194b16. We tested on the following settings:

• python 3.7.7
• cuda 10.1
• pytorch 1.5.0
• torchvision 0.6.0
• mmcv 1.0.4

With settings above, please refer to official guide of mmdetection for installation.

2. Data:

Please download trainset and valset of VisDrone2020-DET dataset and UAVDT-Benchmark-M, then unzip all the files and put them under proper paths.

In order to make better use of mmdetection, please convert the datasets to coco format.

Training

Both training and test commands are exactly the same as mmdetection, so please refer to mmdetection for basic usage.

# Single GPU
python tools/train.py ${CONFIG_FILE}

Please make sure the path of datasets in config file is right.

For example, to train a DSHNet model with Faster R-CNN R50-FPN for trainset of VisDrone:

# Single GPU
python tools/train.py configs/faster_rcnn/vd_faster_rcnn_r101_fpn_tail.py --work-dir checkpoints/vd_faster_rcnn_r101_fpn_tail

Multi-gpu training and test are also supported as mmdetection.

Test

# Single GPU
python tools/tests.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [--out ${RESULT_FILE}] [--eval ${EVAL_METRICS}]

For example (assuming that you have downloaded the corresponding chechkpoints file or train a model by ypurself to proper path), to evaluate the trained DSHNet model with Faster R-CNN R50-FPN for valset of VisDrone:

# Single GPU
python tools/test.py configs/faster_rcnn/vd_faster_rcnn_r101_fpn_tail.py checkpoints/vd_faster_rcnn_r101_fpn_tail/latest.pth --eval bbox

Results and models

Please refer to our paper for complete results.

VisDrone

methods	backbone	map	map50	map75	maps	mapm	mapl	ped.	people	bicycle	car	van	truck	tricycle	awn.	bus	motor	model
FRCNN+FPN+DSHNet	R50	24.6	44.4	24.1	17.5	33.8	36.1	22.5	16.5	10.1	52.8	32.6	22.1	17.5	8.8	39.5	23.7	Google Drive
FRCNN+FPN+DSHNet	R101	24.4	44.3	23.8	17.2	33.6	34.8	21.7	16.0	10.1	52.2	31.6	22.7	17.1	9.5	38.6	24.0	Google Drive
RetinaNet+FPN+DSHNet	R50	16.1	30.2	15.5	9.6	24.0	28.6	14.1	8.9	1.3	48.2	24.8	14.2	8.8	6.0	21.6	13.1	Google Drive
CRCNN+FPN+DSHNet	R50	26.2	45.0	26.3	17.9	36.6	38.9	23.2	16.1	11.2	55.5	33.5	25.2	19.1	10.0	43.0	25.1	Google Drive

uavdt

methods	backbone	map	map50	map75	maps	mapm	mapl	car	truck	bus	model
RetinaNet+FPN+DSHNet	R50	17.8	30.4	19.7	11.9	29.9	27.8	32.1	4.2	17.0	Google Drive

Citation

If you find this useful in your work, please consider citing,

@InProceedings{Yu_2021_WACV,
    author    = {Yu, Weiping and Yang, Taojiannan and Chen, Chen},
    title     = {Towards Resolving the Challenge of Long-Tail Distribution in UAV Images for Object Detection},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2021},
    pages     = {3258-3267}
}