Overview

Highlight

• Automatic labeling of objects without training.

• Support representative UAV object detection datasets, include VisDrone.

TODO

• Release the labeling code.

Data Prepare

Take the VisDrone dataset for an example.

VisDrone to COCO Format

python ./tools/VisDrone2COCO.py 

You will get the dataset folder structure as:

--data
 |-annotations
   |-instances_train2017.json
   |-instances_val2017.json
 |-VisDrone2019-DET-train
   |-images
     |-0000002_00005_d_0000014.jpg
     |-...
 |-VisDrone2019-DET-val
   |-images
     |-0000001_02999_d_0000005.jpg
     |-...

Environment Prepare

conda create -n automatic-labeling python=3.8 -y
conda activate automatic-labeling
pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html
pip install mmengine

pip install -U openmim
mim install "mmcv>=2.0.0"

git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection; pip install -e .; cd ..

pip install git+https://github.com/facebookresearch/segment-anything.git
pip install git+https://github.com/openai/CLIP.git

labeling

First download the following checkpoints and save to './checkpoint'

Checkpoint	Download Link
Faster R-CNN	MMDetection Model
DINO	MMDetection Model
Detic	MMDetection Model
Grounding DINO	Official Model
GLIP	Official Model

Closed-Set Object Detection

Faster R-CNN + SAM

python visdrone_style_eval.py ./data/VisDrone ./configs/faster-rcnn_r50_fpn_2x_coco.py ./checkpoint/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth -t visdrone_cls_name.txt -o ./output/faster-rcnn_r50_fpn_2x_coco -b 0.01

Open Vocabulary Detection

DINO

python visdrone_style_eval.py ./data/VisDrone configs/dino-5scale_swin-l_8xb2-12e_coco.py ./checkpoint/dino-5scale_swin-l_8xb2-12e_coco_20230228_072924-a654145f.pth -t visdrone_cls_name.txt -o ./output/dino-5scale_swin-l_8xb2-12e_coco -b 0.01

Detic + SAM

python visdrone_style_eval.py ./data/VisDrone configs/Detic_LI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.py ./checkpoint/detic_centernet2_swin-b_fpn_4x_lvis-coco-in21k_20230120-0d301978.pth -t visdrone_cls_name.txt -o ./output/Detic_LI21k_CLIP_SwinB_896b32_4x_ft4x_max-size -b 0.01

Grounding Object Detection

GroundingDINO

python visdrone_style_eval.py ./data/VisDrone configs/GroundingDINO_SwinT_OGC.py ./checkpoint/groundingdino_swint_ogc.pth -t visdrone_cls_name.txt -o ./output/GroundingDINO_SwinT_OGC -b 0.01

GLIP + SAM

python visdrone_style_eval.py ./data/VisDrone configs/glip_Swin_L.yaml ./checkpoint/glip_large_model.pth -t visdrone_cls_name.txt -o ./output/glip_large_model -b 0.01

Single Image labeling

python detector_sam_demo.py ./data/VisDrone/VisDrone2019-DET-val/images/0000001_02999_d_0000005.jpg configs/Detic_LI21k_CLIP_SwinB_896b32_4x_ft4x_max-size.py ./checkpoint/detic_centernet2_swin-b_fpn_4x_lvis-coco-in21k_20230120-0d301978.pth -t car