Research map / 3D vision / NeRF and radiance fields
Learn-to-Score: Efficient 3D Scene Exploration by Predicting View Utility
Benjamin Hepp, Debadeepta Dey, Sudipta N. Sinha, Ashish Kapoor, Neel Joshi, Otmar Hilliges
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 |
|---|---|---|---|
| use learned 3D ConvNet utility instead of ProximityCount | next-best-view exploration · Washington2 | efficiency (normalized): 0.52 -> 0.91 | Improved |
| use learned utility instead of the best handcrafted utility | next-best-view exploration · Neighborhood | efficiency (normalized): 0.60 -> 0.74 | Improved |
| use learned utility instead of Frontier | next-best-view exploration · Paris | efficiency (normalized): 0.57 -> 0.87 | Improved |
| use learned utility instead of Frontier | next-best-view exploration · SanFrancisco | efficiency (normalized): 0.09 -> 0.77 | Improved |
| use learned utility instead of ProximityCount, trained on one scene | next-best-view exploration · Washington1 | efficiency (normalized): 0.47 -> 0.88 | Improved |
| replace raycast-based AverageEntropy/ProximityCount scoring with a ConvNet forward pass | utility function · per-step computation | time per step (s): 8.35 -> 0.57 | Improved |
| replace Frontier raycast with learned utility | utility function · per-step computation | time per step (s): 0.61 -> 0.57 | No clear effect |
| apply low sensor noise at test time (40% pixels dropped, sigma 0.1m) | Learn-to-Score · Washington2 | efficiency (normalized): 1.00 -> 0.99 | No clear effect |
| apply very high sensor noise at test time (sigma 1.0m) | Learn-to-Score · Washington2 | efficiency (normalized): 1.00 -> 1.02 | No clear effect |
| use stereo-matched depth images from a photorealistic engine | Learn-to-Score · Washington2 | efficiency (normalized): 1.00 -> 0.99 | No clear effect |
| use learned utility trained on one real scene instead of ProximityCount | next-best-view exploration · outdoor real building | efficiency (normalized): 0.58 -> 0.90 | Improved |
| use learned utility trained on Area1 instead of ProximityCount | next-best-view exploration · indoor 2D-3D-S Area5b | efficiency (normalized): 0.52 -> 0.78 | Improved |
Closest papers by meaning
- MACARONS: Mapping and Coverage Anticipation With RGB Online Self-Supervision (CVPR 2023 · similarity 0.86)
- Learning from Unlabelled Videos Using Contrastive Predictive Neural 3D Mapping (ICLR 2020 · similarity 0.86)
- Where to Explore Next? ExHistCNN for History-aware Autonomous 3D Exploration (ECCV 2020 · similarity 0.85)
- Learning View Selection for 3D Scenes (CVPR 2021 · similarity 0.84)
- Neural Groundplans: Persistent Neural Scene Representations from a Single Image (ICLR 2023 · similarity 0.84)
- Single-View View Synthesis With Multiplane Images (CVPR 2020 · similarity 0.84)
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.
