Omni-Scale Relation Distillation for Object Detection
- 주제(키워드) Knowledge Distillation
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 Wonjun Hwang
- 발행년도 2024
- 학위수여년월 2024. 2
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000033407
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
In the field of object detection, Knowledge Distillation (KD) stands out as a popular compression strategy. Previous KD approaches were primarily concerned with feature imitation in order to consider both classification and localization. However, due to lack of instance relations and the global context in an image, feature-based approaches have difficulties in transferring information for hard examples (e.g., small objects). In this thesis, we present Omni-Scale Relation Distillation, which takes into account global context and instance relation across multi-scale features. We use a clustering-based approach for instance-centric context extraction for global context distillation. To create cluster centroids, we employ key feature selection to eliminate background noise. Furthermore, we suggest dense instance relation distillation among important instance features for emphasizing foreground features. The incorporation of context vectors improves the efficiency of the instance relation distillation process. Extensive experiments with various detectors show that our method is effective, with state-of-the-art performance on the COCO dataset when compared to other distillation approaches. Keywords: Object Detection, Knowledge Distillation, Deep Neural Network, Transfer Learning
more목차
1. Introduction 1
2. Related Work 5
2.1 Object Detection 5
2.2 Knowledge Distillation 6
3. Proposed Methods 8
3.1 Key Feature Selection 8
3.2 Omni-Scale Global Context Distillation 10
3.3 Dense Instance Relation Distillation 12
4. Experiments 14
4.1 Setup and Implementation Details 14
4.2 Main Results 15
4.3 Ablation study 17
5. Conclusion 23
Bibliography 24
