단안 깊이 추정

Self-supervised monocular methods can use adjacent frames and camera motion to reconstruct a target view, minimizing photometric differences. Monodepth2 adds techniques such as per-pixel minimum reprojection and auto-masking. Such methods depend on assumptions about scene motion and camera geometry. Foundation-model releases such as Depth Anything are version-specific; do not infer metric accuracy in a new environment from a benchmark result. A single image provides visual cues such as relative object size, perspective convergence, texture gradients, and occlusion, but not direct stereo disparity. Depth Anything V2’s reported recipe trains a DINOv2-G teacher on synthetic labeled images, uses it to pseudo-label large-scale unlabeled real images, and trains student models on those pseudo-labels. The paper introduces DA-2K, with 1,000 images and 2,000 sparse relative-depth pixel pairs across diverse scenes. Its benchmark results are specific to the paper and do not guarantee performance in every environment or metric scale.