-
Notifications
You must be signed in to change notification settings - Fork 2
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Augmentation] Small object를 위한 Augmentation #52
Comments
resize multiple scale은 A.Oneof에 A.resize 여러개 넣는 식으로 만들면 될까요? |
저는 mmsegmentation에서 기본으로 제공하는 augmentation 중에 Resizemultiscale이었나, 그게 있었던 것 같아서 해당 augmentation을 사용하려고 했습니다. |
transform = []
i = random.randint(0,2)
resize_list = [512, 1024, 768]
transform.append(A.Resize(resize_list[i], resize_list[i]))
# transform.append(A.RandomScale(0.5, p=1.0))
# transform.append(A.OneOf([A.Resize(512, 512), A.Resize(1024, 1024), A.Resize(768, 768), A.Resize(384, 384)], p=1.0)) |
#37 crop을 주는게 좋겠다는 의견을 보고, resize를 하기보단 |
작은 객체를 더 잘 맞추도록 학습하는 것이 mIoU관점에서 좋은 학습 방향이 아닌 것 같다는 생각이 듭니다.
따라서, 성능향상을 위해서 학습 방향에 있어 small object에 집중하는 것은 좋지 않아보입니다. 이에 따라 이슈는 닫아두겠습니다.
활용해보시는 것도 좋아보입니다. 사용 전 사용 후 비교
|
Background
LB에서 0.68정도가 나왔던 파일을 시각화해서 보던 중 아래와 같이 작은 물체의 경우 아예 잘 잡지 못하는 ( receptive field의 문제처럼 느껴졌음 ) 문제를 확인했습니다. 이를 위해서 이번에도 multi-scale전략을 가져가야할 것 같습니다.
이미지는 cocoid기준 16번
Content
이와 같은 맥락에서 augmentation의 관점에서 해볼 수 있는 것들을 해보고 결과를 확인해보자 합니다. 추가적으로 저희가 가지고 있는 데이터의 경우 작은 객체도 확대해서 찍힌 사진이 대부분이어서 확실히 작은 객체들에 대해서 약하다고 느꼈습니다.
Details
The text was updated successfully, but these errors were encountered: