Research map / 3D vision / Depth and stereo
Learning Online Multi-sensor Depth Fusion
Erik Sandström, Martin R. Oswald, Suryansh Kumar, Silvan Weder, Fisher Yu, Cristian Sminchisescu, Luc Van Gool
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 the extra loss terms beyond L_f (full loss) (ablation) | SenFuNet · Replica SGM+PSMNet ablation (no denoising) | F-score (%): 62.65 -> 69.83 | Improved |
| use 2 3D conv layers in weighting network instead of none (ablation) | SenFuNet · Replica SGM+PSMNet ablation (no denoising) | F1 (%): 48.57 -> 69.86 | Improved |
| increase 3D conv layers from 2 to 4 (ablation) | SenFuNet · Replica SGM+PSMNet ablation (no denoising) | F1 (%): 69.86 -> 68.21 | No clear effect |
| add feature network (2D processing before unprojection) (ablation) | SenFuNet · Replica SGM+PSMNet ablation (no denoising) | F-score (%): 67.44 -> 68.87 | Improved |
| add tanh normalization of weight counters (ablation) | SenFuNet · Replica SGM+PSMNet ablation (no denoising) | F-score (%): 66.69 -> 68.92 | Improved |
| add feature grids to weights as weighting-network input (ablation) | SenFuNet · Replica SGM+PSMNet ablation (no denoising) | F-score (%): 68.92 -> 69.83 | Improved |
| use 5 instead of 1 feature-network blocks (ablation) | SenFuNet · Replica SGM+PSMNet ablation (no denoising) | F-score (%): 67.45 -> 68.39 | Mixed |
| vary feature dimension n from 2 to 5 (ablation) | SenFuNet · Replica SGM+PSMNet ablation (no denoising) | F-score (%): 68.86 -> 68.21 | No clear effect |
| add depth denoising network (ablation) | SenFuNet · Replica ToF+PSMNet fusion (with denoising) | F-score (%): 69.29 -> 76.47 | Improved |
| add depth denoising network (ablation) | SenFuNet · Replica SGM+PSMNet fusion (with denoising) | F-score (%): 69.83 -> 71.18 | Improved |
| late 3D fusion SenFuNet versus 2D Early Fusion with synchronized sensors | SenFuNet · Replica ToF+PSMNet fusion (with denoising) | F-score (%): 74.60 -> 76.47 | Improved |
| lower ToF sampling rate from 1/2 to 1/3 | SenFuNet · Replica ToF+PSMNet time-asynchronous (ToF at 1/3 rate) | F-score (%): Early Fusion 61.34 -> 55.52, SenFuNet 73.04 -> 74.93 | 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.
- RoutedFusion — RoutedFusion: Learning Real-Time Depth Map Fusion (CVPR 2020) · better on 4 of 4 shared results
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
- RoutedFusion: Learning Real-Time Depth Map Fusion (CVPR 2020 · similarity 0.90)
- Learning Non-Volumetric Depth Fusion Using Successive Reprojections (CVPR 2019 · similarity 0.88)
- NeuralFusion: Online Depth Fusion in Latent Space (CVPR 2021 · similarity 0.87)
- Robust Depth Estimation Under Sensor Degradations: A Multi-Sensor Fusion Perspective (TPAMI 2025 · similarity 0.87)
- A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · similarity 0.85)
- Robust Depth Enhancement via Polarization Prompt Fusion Tuning (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.
