ν μ΄λ¦ : λ©μμ΄
κΉμ±ν |
λ°μμ |
μ΄λ€ν |
μ΄μ±μ |
μ νΈμ°¬ |
- λν μ£Όμ : μ£Όμ΄μ§ μ¬μ§μμ μ°λ κΈ°λ₯Ό Detectionνλ λͺ¨λΈ ꡬν
- λν λͺ©ν
- 체κ³μ μΈ μ€ν κ΄λ¦¬ (e.g., λΉκ΅λΆμμ μν table μμ±)
- robustν λͺ¨λΈ μ€κ³ (e.g., train/test dataμ λν μ±λ₯μ°¨μ΄κ° μμ λͺ¨λΈ μ€κ³)
- μ κ·Ήμ μΈ GitHub νμ©μ ν΅ν νμ μ§ν (e.g., GitHub flow νμ©)
- λν μΌμ : 23.05.03 ~ 23.05.18 19:00 (2μ£Ό)
μ΄λ¦ | μν | github |
---|---|---|
κΉμ±ν | Detectron2 (cascade, tridentnet, faster rcnn, retinanet) μ€ν, Ensemble | Happy-ryan |
λ°μμ | Detectron2, Torchvision Faster R-CNN μ€ν, Yolo v6 μ€ν, mAP metric λΆμ | nstalways |
μ΄λ€ν | Mmdetection baseline κ΅¬μ± λ° μ€ν, Pseudo labeling/Ensemble μ€ν | DaHyeonnn |
μ΄μ±μ | Mmdetection training baseline κ΅¬μ± λ° μ€ν, λͺ¨λΈ Backbone λ° TTA μ€ν | Chaewon829 |
μ νΈμ°¬ | Detectron2 μ€ν, MMdetection-Cascade Swin L RCNN μ€ν, Augmentation μ€ν | Eumgil98 |
-
μ 체 λ°μ΄ν°μ ν΅κ³
- μ 체 μ΄λ―Έμ§ κ°μ : 9754 μ₯ (train 4883, validation 4871)
- ν΄λμ€ μ’ λ₯ : 10 κ° (General trash, Paper, Paper pack, Metal, Glass, Plastic, Styrofoam, Plastic bag, Battery, Clothing)
- μ΄λ―Έμ§ ν¬κΈ° : (1024, 1024)
-
μ΄λ―Έμ§ μμ
μ μ΄λ―Έμ§λ μμμΌ λΏμ΄λ©°, μ€μ λ°μ΄ν°μλ κ΄λ ¨μ΄ μμ΅λλ€. -
(μ£Όμ) Submission & Annotation format
- Submission formatμ PASCAL VOC νν!
- Annotation formatμ COCO νν!
- formatλ§λ€ bboxλ₯Ό μ μνλ λ°©μμ΄ λ€λ₯΄λ―λ‘, metric κ³μ° μ μ£Όμ!! (Ref)
main
βββ detectron2
β βββ tridentnet : detectron2μμ μ 곡νμ§ μμ λͺ¨λΈ μ¬μ©νκΈ° μν dir
β βββ inference.py : model inference λ° submission file μμ±
β βββ mapper.py : data augmentation λ΄λΉνλ μ½λ
β βββ trainer.py : data loader λ° evaluator μμ±νλ μ½λ
β βββ utils.py : config μ€μ μ½λ
β βββ train.py : νμ΅ μ€ννλ Command Line Interface
β
βββ mmdetection
β βββ augmentation
β β βββ BaseAugmentation.py : bbox annotation load λ° tensor λ³νλ§ ν¬ν¨ν Base Aug
β β βββ CustomAugmentation.py : custonAugmentationμ ꡬμ±νκ³ pipelineμ importνλ μ½λ
β βββ pipeline.py : train, val, testμ Transform pipeline ꡬμ±
β βββ inference.py : model inference λ° submission file μμ±
β βββ train.py : νμ΅ μ€ννλ Command Line Interface
β
βββ torchvision
β βββ configs : train/evaluation/inference μ μ¬μ©νλ yaml νμΌλ€μ λͺ¨μλ ν΄λ
β βββ model : custom model μ½λλ€μ λͺ¨μλ ν΄λ
β βββ trainer : λͺ¨λΈ λ³λ‘ train μ μ¬μ©νλ μ½λλ€μ λͺ¨μλ ν΄λ
β βββ evaluation.py
β βββ inference.py
β βββ my_dataset.py
β βββ my_optimizer.py
β βββ train.py
β βββ transform.py
β βββ utils.py
β
βββ yolov6
βββ custom_dataset.py : yolov6μμ μꡬνλ νμμ λ§κ² λλ ν 리λ₯Ό μ¬κ΅¬μ±νλ μ½λ
βββ recycle.yaml : λ°μ΄ν°μ
μ κ²½λ‘ λ° classμ λν μ λ³΄κ° λ΄κ²¨μλ yaml νμΌ
βββ submission.py
π
Private score : 9 / 19
π
Public score : 9 / 19
Model
βββ 2 Stage Model
β βββ Faster RCNN :0.5385
β βββ Cascade RCNN :0.5747
β βββ DETR : 0.3987
βββ 1 Stage Model
βββ PAA : 0.5787
βββ UniverseNet :0.6383
βββ RetinaNet : 0.3406
βββ TOOD : 0.4482
βββ YOLOv6 :0.5424