Research project associated with IEEE Robotics and Automation Letters (RA-L).
Core approach
Combines geometric consensus-aware cost aggregation with geometrically anchored feature rectification to improve multi-view depth estimation.
GC-CA suppresses depth hypotheses that disagree with multi-view geometric consensus. GAFR-Net uses the multi-view cost volume as a metric anchor to rectify monocular features. Evaluation covers KITTI, 7-Scenes, and zero-shot transfer to TUM-RGBD.
Associated publication
GeoFusion: Geometric Consensus Learning and Anchored Feature Fusion for Multi-View Depth Estimation