boostcampaitech3 / level2-object-detection-level2-cv-08

level2-object-detection-level2-cv-08 created by GitHub Classroom
1 stars 3 forks source link

Use Cascade RCNN MulitScale #3

Open hursuk1 opened 2 years ago

hursuk1 commented 2 years ago

세부 설명

Cascade RCNN 구조 사용 Backbone : SwinTransformer large patch4_window7_224

ms train : (512, 512), (608, 608), (704, 704), (736, 736), (832, 832), (864, 864), (960, 960), (992, 992), (1056, 1056), (1088, 1088), (1152, 1152), (1184, 1184), (1248, 1248), (1280, 1280), ms test : [(512, 512), (1024, 1024), (1536, 1536)]

mAP epoch 9 : 0.6074 epoch 17 : 0.6279

추론 : 이미지 분류와 다르게 객체 검출은 학습을 오래해서 학습량이 많을 수록 좋은 성능을 보인다고 생각됨