Observability-Aware Rotation-Target Filtering for Self-Supervised Depth-Only Odometry
Main Article Content
Abstract
Plane-based pseudo-labels can provide complete pose targets even when scene geometry supports only some motion components. We introduce an observabilityaware training approach that gates rotation targets using matched-plane normal separation and projects translation residuals onto separately supported directions. The student is a depth-only model at inference and receives no geometric support indicators. We compare it with unfiltered supervision using the same 620,169- parameter student, five paired random seeds, equal training budgets, and a common data protocol. The evaluation covers nine sequences from TUM RGB-D, VCURVI and ETH3D, including four ETH3D zero-shot transfer sequences. Across all nine sequences, the filtered model has lower mean rotation root-mean-square error (RMSE) but higher mean translation RMSE than the unfiltered model. Thus the benefit is component-specific rather than a uniform pose improvement. Finite contiguoustrajectory tests likewise show no uniform trajectory benefit and reveal unfavorable translation accumulation. TUM and ETH3D provide verified-camera physical evidence; VCU results remain conditional because the supplied reference-frame semantics are not independently verified. The method is therefore best understood as targeted supervision for rotation under geometric degeneracy, not as a guarantee of globally improved odometry.