Research map / 3D vision / Depth and stereo
Deep Cost Ray Fusion for Sparse Depth Video Completion
Jungeon Kim, Soongjin Kim, Jaesik Park, Seungyong Lee
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ray-based cost volume fusion (self+cross attention) (ablation) | RayFusion base network · KITTI val (20% train) | MAE (mm): 203.05 -> 198.51 | Improved |
| use convolutional GRU local fusion instead of ray-based fusion (ablation) | RayFusion · KITTI val (20% train) | MAE (mm): 215.18 -> 228.85 | Hurt |
| add L1 depth loss to cross-entropy loss (ablation) | RayFusion base network · KITTI val (20% train) | MAE (mm): 223.59 -> 203.05 | Improved |
| add L2 loss to CE + L1 (ablation) | RayFusion base network · KITTI val (20% train) | MAE (mm): 203.05 -> 220.22 | Hurt |
| add NLSPN-based depth refinement module (ablation) | RayFusion · KITTI val | MAE (mm): 182.65 -> 176.23 | Improved |
| add NLSPN-based depth refinement module (ablation) | RayFusion · ScanNetV2 test | MAE (m): 0.0159 -> 0.0160 | No clear effect |
| add NLSPN-based depth refinement module (ablation) | RayFusion · VOID test (0.5%) | MAE (mm): 24.57 -> 24.51 | No clear effect |
| add temporal cross-attention fusion vs single-view setting (ablation) | RayFusion · ScanNetV2 test | MAE (m): 0.0208 -> 0.0159 | Improved |
| ray-based cost volume fusion vs prior methods on unseen sparsity | RayFusion · VOID (train 0.5%, test 0.05%) | MAE (mm): 104.97 -> 78.55 | Improved |
| RayFusion cross-dataset transfer vs CompletionFormer | RayFusion · ScanNetV2 train -> VOID test | MAE (mm): 65.90 -> 29.08 | Improved |
| use soft labels instead of one-hot for cost volume supervision (ablation) | RayFusion · depth completion | performance | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SimpleRecon — SimpleRecon: 3D Reconstruction without 3D Convolutions (ECCV 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Sparse-to-Dense Depth Completion Revisited: Sampling Strategy and Graph Construction (ECCV 2020 · similarity 0.89)
- Improving Depth Completion via Depth Feature Upsampling (CVPR 2024 · similarity 0.86)
- RaySt3R: Predicting Novel Depth Maps for Zero-Shot Object Completion (NeurIPS 2025 · similarity 0.86)
- DeCoTR: Enhancing Depth Completion with 2D and 3D Attentions (CVPR 2024 · similarity 0.86)
- Zero-shot Depth Completion via Test-time Alignment with Affine-invariant Depth Prior (AAAI 2025 · similarity 0.85)
- Completion as Enhancement: A Degradation-Aware Selective Image Guided Network for Depth Completion (CVPR 2025 · 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.
