Boostcamp Recycle Trash Object Detection Challenge

Code for 4th place solution in Boostcamp AI Tech Recycle Trash Object detection Challenge.

대량 생산, 대량 소비의 시대에 살며 '쓰레기 대란', '매립지 부족'과 같은 여러 사회 문제를 낳고 있다.
분리수거는 이러한 환경부담을 줄일 수 있는 방법이다. 해당 대회는 쓰레기를 detection하는 모델을 만들어 정확한 분리수거를 돕는 것에 기여한다.

Contributors
김서원_T2036, 이유진_T2167, 이한빈_T2176, 정세종_T2201, 조현동_T2215, 허지훈_T2241, 허정훈_T2240

Archive contents

detection
├── dataset
├── template
│   ├──mmdetection
│   │  ├──configs
│   │  │  └──custom
│   │  │     ├──helper
│   │  │     │  ├──dateset.py
│   │  │     │  ├──runtime.py
│   │  │     │  └──schedule.py
│   │  │     └──models
│   │  │        ├──cascade_rcnn
│   │  │        ├──faster_rcnn
│   │  │        └──htc
│   │  ├──tools
│   │  │  ├──train.py
│   │  │  ├──inference.py
│   │  │  ├──ensemble.py
│   │  │  ├──make_fold_annotation.py
│   │  │  └──vis_submission.ipynb
│   │  └──submission
│   │     ├──ensemble_inference.py
│   │     └──ensemble_inf_cfg.json
│   ├──live
└── └──detectron

Train

cd mmdetection

1. vanilla train

python tools/train.py [config path]

2. k-fold train

python tools/make_fold_annotation.py [original_train_json_path]
python tools/train_cv.py [config path]

3. pseudo labeling train & inference

python tools/inference_with_pseudo_labeling.py [config path]

4. optimization with wandb sweeps

• Setting
  • change sweep.yaml file
• Command

  • create sweep graph

    wandb sweep sweep.yaml
    

  • then you can get url

  • you have to change sweepID

    wandb agent ProjectName/sweepID
    

inference

cd mmdetection

1. vanilla inference

python tools/inference.py [config path]

2. k-fold inference

python tools/inference_cv.py [config names] [work_dir]

visualization result

• Make submission csv file after training
• Change PRED_CSV in vis_submission.ipynb
• Run cells

ensemble

cd submission

1. Modify ensemble_inf_cfg.json
2. Run ensemble_inference.py

python ensemble_inference.py ensemble_inf_cfg.json

yolo

https://github.com/ultralytics/yolov5

EfficeintDet

https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch

final ensemble model

ensemble_method : Weighted Boxes Fusion

Model and Experiments

• "htc_swin_b_384.csv"
• "hsbfm_with_psudolabilng_8.csv"
• "htc_pvt_finetune_mosaic_final.csv"
• "htc_swin_b_finetune_mosaic.csv"
• "htc_swin_b.csv"
• "htc_swin_b_kfold.csv"
• "cascadercnn_pvt.csv"
• "faster_rcnn_pvtv2_b5_final.csv"

Model(detector)	Exepriments Result
Hybrid Task Cascade	README
CasCade R-CNN	README
Faster R-CNN	READEM