Infrared Monocular Depth Estimation via Parameter-Efficient Adaptation

초록

Recent RGB-based monocular depth estimation achieves high accuracy but often assumes stable visible illumination. Infrared radiation (IR) imaging is practical because it remains usable in dim environments and employs invisible active illumination, which does not interfere with human vision, making it suitable for everyday deployment. However, paired IR-depth datasets are scarce, which limits both training and fair evaluation for this modality. We adapt the large pre-trained model Depth Anything v2 to the IR modality using parameter-efficient LoRA fine-tuning. To support training and evaluation, we collect an indoor dataset of $25,743$ paired IR-depth frames from 7 subjects. Our results show that LoRA-based adaptation transfers effectively to IR without full retraining, delivering accurate depth predictions while keeping memory and compute overhead modest. This study provides a practical pipeline for IR-based depth estimation and highlights the value of modality-aligned data and lightweight adaptation for real-world systems. © 2026 IEEE.

키워드