Adaptive Hybrid Patch Sampling for Deep Patch Visual Odometry

Deep patch-based visual odometry estimates camera trajectories by tracking a compact set of local image patches, offering a favorable balance between accuracy and efficiency. However, patch quality remains critical under a fixed patch budget. Random sampling preserves spatial diversity but may select weakly informative regions, while response-driven sampling tends to cluster patches around salient structures and reduce spatial coverage. To address this issue, we propose Adaptive Hybrid Patch Sampling (AHPS), a lightweight front-end sampler for deep patch visual odometry. AHPS first predicts patch utility through local feature aggregation and a shared voting layer, then combines Top-K score-guided selection with random sampling from the remaining positions. Without modifying the subsequent patch tracking and bundle adjustment backend, AHPS improves both patch discriminability and spatial coverage. Experiments on ICL-NUIM, TartanAir, and TUM RGB-D show that AHPS consistently improves trajectory accuracy over the DPVO baseline.