Research map / 3D vision / Depth and stereo
DELTAS: Depth Estimation by Learning Triangulation And densification of Sparse points
Ayan Sinha, Zak Murez, James Bartolozzi, Vijay Badrinarayanan, Andrew Rabinovich
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 learned triangulation module and detector/descriptor heads in front of the sparse-to-dense network (ablation) | sparse-to-dense depth network · ScanNet test (3 frames, ablation) | Abs Rel: 0.1242 -> 0.0932 | Improved |
| add triangulated sparse points (vs none) (ablation) | DELTAS · ScanNet test (3 frames, ablation) | Abs Rel: 0.2203 -> 0.0932 | Improved |
| use 32 triangulated points instead of none (ablation) | DELTAS · ScanNet test (3 frames, ablation) | Abs Rel: 0.2203 -> 0.1105 | Improved |
| increase points from 512 to 768 (ablation) | DELTAS · ScanNet test (3 frames, ablation) | Abs Rel: 0.0932 -> 0.0942 | No clear effect |
| swap triangulated depth for ground-truth depth at the 512 points (ablation) | DELTAS · ScanNet test (3 frames, ablation) | Abs Rel: 0.0932 -> 0.0680 | Improved |
| raise NMS radius from 3 to 9 (ablation) | DELTAS · ScanNet test (3 frames, ablation) | Abs Rel: 0.0942 -> 0.0932 | Improved |
| raise NMS radius from 9 to 11 (ablation) | DELTAS · ScanNet test (3 frames, ablation) | Abs Rel: 0.0932 -> 0.0938 | Hurt |
| use 2 pixel offset instead of 1 pixel (ablation) | DELTAS · ScanNet test (3 frames, ablation) | Abs Rel: 0.0932 -> 0.0933 | No clear effect |
| vary ratio of detector points to random points (ablation) | DELTAS · ScanNet test (3 frames, ablation) | Abs Rel: 0.0935 -> 0.0932 | No clear effect |
| vary descriptor length 25 to 150 (ablation) | DELTAS · ScanNet test (3 frames, ablation) | Abs Rel: 0.0934 -> 0.0932 | No clear effect |
| replace ResNet-50 backbone with VGG-9 (ablation) | DELTAS · ScanNet test (3 frames, ablation) | Abs Rel: 0.0932 -> 0.1073 | Hurt |
| add more input views from 2 to 7 frames | DELTAS · ScanNet test (n frames) | Abs Rel: 0.106 -> 0.087 | Improved |
Later papers that compared with this method
3 later accepted papers report DELTAS as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- SimpleRecon: 3D Reconstruction without 3D Convolutions (ECCV 2022 · better on 2 of 2 shared results)
- A Confidence-Based Iterative Solver of Depths and Surface Normals for Deep Multi-View Stereo (ICCV 2021 · better on 4 of 4 shared results)
- NerfingMVS: Guided Optimization of Neural Radiance Fields for Indoor Multi-View Stereo (ICCV 2021 · better on 2 of 2 shared results)
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.
- DPSNet — DPSNet: End-to-end Deep Plane Sweep Stereo (ICLR 2019) · better on 5 of 6 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
- MVS2D: Efficient Multi-View Stereo via Attention-Driven 2D Convolutions (CVPR 2022 · similarity 0.90)
- Fast-MVSNet: Sparse-to-Dense Multi-View Stereo With Learned Propagation and Gauss-Newton Refinement (CVPR 2020 · similarity 0.90)
- Generalized Binary Search Network for Highly-Efficient Multi-View Stereo (CVPR 2022 · similarity 0.90)
- A Survey on Deep Learning Techniques for Stereo-Based Depth Estimation (TPAMI 2020 · similarity 0.90)
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.89)
- LEAP: Liberate Sparse-View 3D Modeling from Camera Poses (ICLR 2024 · similarity 0.88)
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.
