Efficient Domain Adaptation for Event-Based Monocular Depth Estimation

다국어 초록 (Multilingual Abstract)

This thesis examines efficient transfer of an event-based monocular depth estimator when the source data and the deployment data are captured by different real event-camera systems. Event sensors do not acquire dense frames; instead, they asynchronously report pixel-level brightness changes. This sensing mode is useful for rapid motion and difficult illumination, but the resulting event stream is sparse, irregular, and strongly dependent on the sensor and scene. As a result, a model trained on one event-camera benchmark can lose accuracy when it is applied to another benchmark. The study considers the adaptation of a recurrent event-depth backbone trained on MVSEC to the DSEC driving domain. The main question is how much target-domain accuracy can be recovered when the base network is kept almost entirely fixed and only a very small set of parameters is optimized using a limited labeled calibration subset. The experiments are organized around RDANet, a recurrent adaptation framework that uses a Depth AnyEvent-R-style backbone with a DINOv2 visual-foundation encoder, ConvLSTM temporal processing, and a dense depth decoder. The thesis evaluates several parameter-efficient fine-tuning methods under the same MVSEC-to-DSEC setting. The compared methods are Bias-only tuning, adapter tuning, decoder- weight tuning, ConvLSTM-only tuning, LoRA at ranks r = 4, 8, and 16, and feature-wise linear modulation (FiLM). For fairness, all methods share the same MVSEC source checkpoint, DSEC calibration budgets, DSEC validation split, and three-seed evaluation protocol. The experimental results show that compact parameter updates can substantially reduce the cross-dataset gap. Bias-only tuning gives the most reliable performance in the N = 150 calibration setting, LoRA with r = 4 is competitive when only N = 100 calibration samples are available, and larger LoRA ranks are less stable under the limited-data protocol. Decoder-side FiLM gives the smallest adaptation footprint while remaining close to the larger PEFT alternatives. These findings support a deployment strategy that combines a frozen recurrent event-depth backbone with small domain-specific adaptation parameters.

목차 (Table of Contents)

• CHAPTER 1. INTRODUCTION 11
• 1.1 Background 11
• 1.2 Research Gap and Motivation 12
• 1.3 Research Objectives 12
• 1.4 Contributions 13
• 1.5 Thesis Organization 13
• CHAPTER 2. RELATED WORK 14
• 2.1 Image-Based Monocular Depth Estimation 14
• 2.2 Event-Based Vision and Event Representations 14
• 2.3 Event-Based Depth Estimation 14
• 2.4 Domain Adaptation for Depth Estimation 15
• 2.5 Parameter-Efficient Fine-Tuning 15
• 2.6 Position of This Thesis 16
• CHAPTER 3. DATASETS AND PROBLEM FORMULATION 17
• 3.1 Overview of the Transfer Setting 17
• 3.2 MVSEC Source Dataset 17
• 3.3 DSEC Target Dataset 17
• 3.4 Cross-Dataset Shift 18
• 3.5 Formal Problem Definition 18
• 3.6 Evaluation Metrics 18
• CHAPTER 4. PROPOSED METHOD 20
• 4.1 RDANet Overview 20
• 4.2 Event Encoding 20
• 4.3 Frozen Recurrent Depth Backbone 21
• 4.4 FiLM-Based Adaptation 21
• 4.5 Alternative PEFT Baselines 22
• 4.6 Optimization Strategy 23
• CHAPTER 5. EXPERIMENTAL SETUP 24
• 5.1 Source Training and Target Calibration 24
• 5.2 Implementation Details 24
• 5.3 Evaluation Protocol 24
• 5.4 Computational Cost Measurement 25
• CHAPTER 6. RESULTS AND ANALYSIS 26
• 6.1 In-Domain MVSEC Evaluation 26
• 6.2 Cross-Domain DSEC Evaluation 26
• 6.3 Parameter-Efficient Adaptation Comparison 28
• 6.4 Statistical Significance 28
• 6.5 FiLM Placement Study 29
• 6.6 Qualitative Analysis 30
• CHAPTER 7. DISCUSSION 31
• 7.1 Interpretation of Bias-Only Adaptation 31
• 7.2 Why Decoder-Side FiLM Is Useful 31
• 7.3 Deployment Implications 31
• 7.4 Limitations 31
• 7.5 Future Research Directions 32
• CHAPTER 8. CONCLUSION 33
• REFERENCES 34
• Abstract in Korean 36