티스토리 뷰
[Deep Learning] 관심이 생긴 Detection 모델 Plain-DETR, DETR Does Not Need Multi-Scale or Locality Design
developer0hye 2023. 12. 6. 18:25https://github.com/impiga/Plain-DETR
GitHub - impiga/Plain-DETR: [ICCV2023] DETR Doesn’t Need Multi-Scale or Locality Design
[ICCV2023] DETR Doesn’t Need Multi-Scale or Locality Design - GitHub - impiga/Plain-DETR: [ICCV2023] DETR Doesn’t Need Multi-Scale or Locality Design
github.com
ICCV 2023 Open Access Repository
DETR Does Not Need Multi-Scale or Locality Design Yutong Lin, Yuhui Yuan, Zheng Zhang, Chen Li, Nanning Zheng, Han Hu; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 6545-6554 Abstract This paper presents an impro
openaccess.thecvf.com
이미 multi-scale feature pyramid 라는 Detection 분야의 하나의 Inductive bias가 된 기법을 걷어내려는 연구는 존재했고, 리뷰도 했었다.
https://developer0hye.tistory.com/333
Exploring Plain Vision Transformer Backbones for Object Detection 리뷰
https://arxiv.org/abs/2203.16527 Exploring Plain Vision Transformer Backbones for Object Detection We explore the plain, non-hierarchical Vision Transformer (ViT) as a backbone network for object detection. This design enables the original ViT architecture
developer0hye.tistory.com
약간 위 논문과 비슷한 종류의 연구같다.
대신 여기서는 Backbone으로 아예 Swin 을 사용했다. 좀 비슷한 점은 "MIM(Masked Image Modeling)이 Detection Task에 있어 Pre-training에 있어서 정말 중요하다."라는 점을 한 번 더 밝혀냈고, 개선된 DETR 방법들로부터 비교적 간단하면서도 성능에 효과적인 방법들에 대한 ablation study를 잘해줬고, 또 Positional Embedding과 Regression 부분의 스케일을 조정함으로써 높은 성능을 달성함을 보이고 있다.

코드와 논문을 대강 보았을때는 Deformable Attention 도 딱히 안 쓴 거 같고, Label assignment 알고리즘도 크게 복잡해보이지 않는다.
드디어, Transformer를 Base로 하면서 과도하게(?) 복잡하지 않은 구조의 Detection 모델이 공개된 거 같다.
'Deep Learning' 카테고리의 다른 글
YOLOv8 low-level output visualization code (0) | 2024.02.02 |
---|---|
Apple이 Vision Transformer 를 자사의 NPU에서 효율적으로 구동시키기 위해 시도한 것들 (0) | 2024.01.22 |
DETR 이 수렴 속도가 느린 의심되는 이유 중 하나 (2) | 2023.12.02 |
Wandb 강제 종료(Ctrl+C) 후 잦은 에러 메시지 출력 문제 해결법 (3) | 2023.11.08 |
[Deep Learning] 20231105 학습시 배치 사이즈에 관하여... (0) | 2023.11.05 |
- Total
- Today
- Yesterday
- 파이참
- Lowest Common Ancestor
- 위상 정렬 알고리즘
- 백준 11437
- 가장 긴 증가하는 부분 수열
- cosine
- PyCharm
- LCA
- FairMOT
- 백준 1766
- 백준
- 백트래킹
- 조합
- 순열
- 단축키
- 인공지능을 위한 선형대수
- MOT
- 자료구조
- ㅂ
- 이분탐색
- 백준 11053
- 문제집
- C++ Deploy
일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
1 | 2 | 3 | 4 | 5 | ||
6 | 7 | 8 | 9 | 10 | 11 | 12 |
13 | 14 | 15 | 16 | 17 | 18 | 19 |
20 | 21 | 22 | 23 | 24 | 25 | 26 |
27 | 28 | 29 | 30 |