Depth Estimation

Depth estimation studies how to recover per-pixel distance and scene geometry from images, video, stereo pairs, and sensors such as LiDAR, radar, event, thermal, and time-of-flight cameras. Methods range from stereo and multi-view geometry to self-supervised monocular learning, sparse-to-dense depth completion, and pretrained vision or video models guided by sparse metric measurements. Current systems increasingly use temporal consistency, uncertainty estimates, camera-aware adaptation, and sensor fusion in settings such as driving, endoscopy, panoramas, and transparent scenes. A central challenge is generalizing across cameras, domains, lighting, and motion while balancing metric accuracy, geometric consistency, and real-time efficiency; benchmarks such as KITTI and NYUv2 also highlight the limits of evaluation when reliable ground-truth depth is scarce.