講演情報
[B-15-52]Boundary-Aware Video Depth Estimation for Far-Field Object Preservation
〇Xinyu LI1, Cedric CAREMEL1, Shuyue FENG1, Yoshihiro KAWAHARA1 (1. The University of Tokyo)
キーワード：
Video Depth Estimation、Far-Field Object Preservation、Boundary-Aware Supervision、3D Reconstruction
Video depth estimation is essential for 3D reconstruction, view synthesis, and immersive display. Although recent pretrained video depth models achieve strong global accuracy and temporal consistency, they still struggle to preserve distant objects. Far-field objects often occupy small image regions and have weak depth discontinuities, causing them to be over-smoothed or merged into the background.
This study investigates a boundary-aware training strategy for far-field object preservation. Without modifying the inference architecture, we add structural supervision to a pretrained video depth framework. The strategy combines spatial consistency, temporal consistency, and boundary-aware refinement to strengthen learning around object-background transitions, especially in distant regions.
Preliminary experiments on indoor and outdoor benchmarks show clearer far-field boundaries and more stable object-background separation than the baseline. Quantitative analysis in distant regions indicates about an 11% reduction in AbsRel and RMSE. These results suggest that boundary-aware supervision improves far-field depth quality while maintaining global depth fidelity.
This study investigates a boundary-aware training strategy for far-field object preservation. Without modifying the inference architecture, we add structural supervision to a pretrained video depth framework. The strategy combines spatial consistency, temporal consistency, and boundary-aware refinement to strengthen learning around object-background transitions, especially in distant regions.
Preliminary experiments on indoor and outdoor benchmarks show clearer far-field boundaries and more stable object-background separation than the baseline. Quantitative analysis in distant regions indicates about an 11% reduction in AbsRel and RMSE. These results suggest that boundary-aware supervision improves far-field depth quality while maintaining global depth fidelity.