Rotation-invariant self-supervised learning with p4-equivariant CNNs = p4-등변 CNN을 활용한 회전 불변 자기지도학습

다국어 초록 (Multilingual Abstract)

This dissertation proposes Guiding Invariance with Equivariance (GIE), a self-supervised framework that achieves rotation-robust representation learning by guiding invariance through group-equivariant modeling. Unlike conventional approaches that enforce invariance via data augmentation, GIE learns it by interpreting structured transformations within an exact p4-equivariant CNN backbone. A learnable orientation predictor and a guided alignment mechanism transform equivariant features into invariant embeddings, producing representations that remain discriminative under rotation. Experiments on STL-10, ImageNet-100, MTARSI, and Pascal VOC demonstrate superior rotation robustness and generalization over prior methods. GIE establishes a unified and efficient approach to geometry-aware self-supervised learning.

국문 초록 (Abstract)

본 논문에서는 회전 변환에 강인한 표현 학습을 위해 등변성을 활용하여 불변성을 유도하는 자기지도학습 프레임워크인 GIE(Guiding Invariance with Equivariance)를 제안한다. 기존 방법들이 데이터 증강을 통해 불변성을 강제하는 것과 달리, GIE는 정확한 p4-등변 합성곱 신경망 기반에서 변환의 구조를 학습적으로 해석함으로써 불변 표현을 형성한다. 학습 가능한 방향 예측 모듈과 정렬 메커니즘을 통해 등변적 특징을 회전 불변 임베딩으로 변환하며, 이를 통해 회전 변화에도 판별력이 유지되는 표현을 학습한다. STL-10, ImageNet-100, MTARSI, Pascal VOC 데이터셋에서의 실험 결과, 제안한 방법은 기존 기법보다 우수한 회전 강인성과 일반화 성능을 달성하였다. GIE는 기하학적 구조를 이해하는 자기지도학습의 통합적 접근법을 제시한다.

목차 (Table of Contents)

• Chapter 1. Introduction 1
• 1.1. Research Motivation 1
• 1.2. Problem Definition and Challenges 3
• 1.3. Research Objectives and Contributions 4
• 1.4. Dissertation Structure 5
• Chapter 2. Related Works 6
• Chapter 3. Theoretical Background 10
• 3.1. Self-Supervised Representation Learning 10
• 3.2. Group Equivariance and p4-Equivariant CNNs 14
• 3.2.1. Group-Equivariant Convolution 14
• 3.2.2. p4-Equivariant CNNs 16
• 3.2.3. Exact Equivariance Condition 17
• Chapter 4. Proposed Method: Guiding Invariance with Equivariance 19
• 4.1. Overview 20
• 4.2. Exact p4-Equivariant Backbone 22
• 4.3. Predicting Equivariance: Orientation Scoring 24
• 4.4. Guiding Invariance: Feature Alignment Mechanism 29
• 4.5. Training Objective 33
• 4.6. Theoretical Analysis 34
• Chapter 5. Experimental Evaluation 37
• 5.1. Experimental Setup 37
• 5.1.1. Datasets 37
• 5.1.2. Training Configuration 39
• 5.1.3. Evaluation Protocols 40
• 5.2. Main Results 41
• 5.2.1. Comparison between p4-CNN Baselines and GIE 42
• 5.2.2. Comparison with Standard CNNs 43
• 5.2.3. Discussion of Performance Trends 43
• 5.3. Ablation Studies and Analysis 45
• 5.3.1. Orientation Predictor and Alignment Loss 45
• 5.3.2. Analysis of Equivariance Scores 46
• 5.3.3. Hyperparameter Sensitivity: β and Hidden Dimension 50
• 5.3.4. Backbone Scale and Parameter Efficiency 52
• 5.3.5. Necessity of Exact Equivariance 53
• 5.4. Generalization and Robustness 55
• 5.4.1. Generalization to Arbitrary Rotation Angles 56
• 5.4.2. Application to Real-World Rotation-Agnostic Domains 58
• 5.4.3. Cross-Dataset Transfer to Non-Pretrained Domains 59
• 5.4.4. Transfer to Semantic Segmentation 59
• Chapter 6. Limitations and Future Work 62
• 6.1. Limitations 62
• 6.2. Future Work 64
• Abstract (in Korean) 72
• Acknowledgement (in Korean) 73