Research map / Point clouds / Point cloud classification
Tri-Perspective View Decomposition for Geometry-Aware Depth Completion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TPVD, which decomposes a sparse LiDAR/ToF point cloud into top, front and side 2D views, fuses them with recurrent 2D-3D-2D TPV Fusion using distance-aware spherical convolution, and refines with a geometric spatial propagation network, and releases the TOFDC smartphone depth dataset.
Problem
Depth completion for autonomous driving and indoor scenes: 2D methods lose 3D geometry, while existing 2D-3D methods use raw point clouds that are too sparse and unevenly distributed to capture fine-grained geometry.
Approach
Decompose the 3D point cloud into three orthogonal 2D views (top, front, side) so the sparse depth stays in 2D and can be densified with 2D convolutions, while recurrent 2D-3D-2D fusion injects 3D geometry via distance-aware spherical convolution and a GSPN builds affinity neighbors across the three TPV spaces and their joint 3D projection.
Key results
- On the KITTI online leaderboard TPVD reaches RMSE 693.97 mm and MAE 188.60 mm, ranking first on all four metrics at submission time.
- On NYUv2 TPVD achieves RMSE 0.086 m and REL 0.010, the best among compared methods.
- On TOFDC TPVD reduces RMSE by 15.6% (0.092 m vs 0.109 m for PointDC) and REL by 33.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TPVD instead of BEV@DC | depth completion · KITTI online test | RMSE (mm): 697.44 -> 693.97 | Helped |
| use TPVD instead of BEV@DC | depth completion · NYUv2 test | RMSE (m): 0.089 -> 0.086 | Helped |
| use TPVD instead of PointDC | depth completion · TOFDC evaluation | RMSE (m): 0.109 -> 0.092 | Helped |
| add top-view depth (ablation) | TPVD front-view baseline · KITTI validation split | RMSE (mm): 763.56 -> 755.14 | Helped |
| add side-view depth (ablation) | TPVD with front and top views · KITTI validation split | RMSE (mm): 755.14 -> 749.38 | Helped |
| add DASC (ablation) | TPVD with three views · KITTI validation split | RMSE (mm): 749.38 -> 735.57 | Helped |
| add GSPN (ablation) | TPVD with DASC · KITTI validation split | RMSE (mm): 735.57 -> 718.90 | Helped |
| increase GSPN neighbors from 5 to 9 (ablation) | GSPN · NYUv2 test | RMSE (m): 3.3 mm | Helped |
| increase GSPN propagation iterations (ablation) | GSPN · NYUv2 test | RMSE (m) | Helped |
| increase TPV Fusion recurrent steps beyond 4 (ablation) | TPV Fusion · KITTI validation split | error | No clear effect |
Limitations
- Parameter count (31.2M) is higher than some lightweight 2D-3D baselines such as ACMNet (4.9M)
- Gains on KITTI validation are modest in some ablation steps, and the recurrent fusion gains saturate after 4 steps
Experimental setup
Datasets: KITTI NYUv2 SUN RGBD TOFDC VKITTI
Benchmarks: KITTI depth completion leaderboard NYUv2 TOFDC KITTI validation split SUN RGBD
Metrics: RMSE MAE iRMSE iMAE REL delta1 delta2 delta3
Baselines: CSPN NLSPN ACMNet RigNet BEV@DC PointDC CFormer LRRU
Builds on: CSPN NLSPN PointDC BEV@DC ACMNet
Closest papers by meaning
- Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction (CVPR 2023 · similarity 0.88)
- DeCoTR: Enhancing Depth Completion with 2D and 3D Attentions (CVPR 2024 · similarity 0.87)
- Three Cars Approaching within 100m! Enhancing Distant Geometry by Tri-Axis Voxel Scanning for Camera-based Semantic Scene Completion (CVPR 2025 · similarity 0.87)
- A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · similarity 0.85)
- RaySt3R: Predicting Novel Depth Maps for Zero-Shot Object Completion (NeurIPS 2025 · similarity 0.85)
- Adaptive Fusion of Single-View and Multi-View Depth for Autonomous Driving (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.