Metric-consistent road-surface novel view synthesis under road-centric sparse-view acquisition
Mochuan Zhan, Terence Patrick Morley, Martin J. L. Turner · Computers & Graphics · 2026
Road-surface-centred novel-view synthesis remains challenging under vehicle-mounted acquisition. Near-linear forward motion and grazing-angle observations leave road geometry weakly constrained. Sparse or repetitive texture, reflections, and far-range perspective compression further degrade image correspondence, stereo depth, and photometric optimisation. We formulate this setting as a metric-consistency problem and introduce a three-phase stereo–LiDAR neural rendering framework that keeps metric information explicit from pose estimation to ray-level supervision. The first stage builds a SuperPoint–SuperGlue stereo-temporal graph that combines temporal tracking edges with fixed-baseline stereo edges to improve SfM coverage under forward road motion. The second stage converts RAFT-Stereo disparity into a sensor-supported dense Z-depth prior through online LiDAR-based scale estimation and log-domain residual propagation. The final stage trains Instant-NGP with anchor-distance confidence-weighted structural losses, so completed-depth pixels close to direct LiDAR support impose stronger ray-level constraints than weakly supported regions. On RSRD (Zhao et al., 2023), the pose front end produces full-sequence reconstructions for all 15 evaluated sequences. Under a near-view held-out-frame protocol over six scenes, the confidence-aware renderer achieves 30.25 dB mean PSNR and 0.7883 mean SSIM, and reduces rendered-depth MAE at held-out LiDAR pixels from 1.81 m with RGB-only training to 0.48 m. The depth prior reduces held-out LiDAR sample AbsRel from 0.0141 for RAFT-Stereo with only per-frame metric conversion to 0.0088 after log-domain calibration under a 10% anchor/90% held-out protocol. This held-out LiDAR evaluation measures metric consistency within the same stereo–LiDAR system rather than independent dense-depth accuracy. Together, the results show improved near-view image quality and rendered-depth consistency under road-centric stereo–LiDAR acquisition.