Research map / 3D vision / 3D generation
3D-RelNet: Joint Object and Relational Network for 3D Prediction
Nilesh Kulkarni, Ishan Misra, Shubham Tulsiani, Abhinav Gupta
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 |
|---|---|---|---|
| combine unary and relative predictions at both train and test time (ablation) | multi-task model · SUNCG (ground-truth boxes, combination ablation) | translation accuracy %: 81.9 -> 84.0 | Improved |
| combine predictions at test time only (ablation) | multi-task model · SUNCG (ground-truth boxes, combination ablation) | translation accuracy %: 81.9 -> 83.5 | Improved |
| combine predictions at train time only (ablation) | multi-task model · SUNCG (ground-truth boxes, combination ablation) | translation accuracy %: 81.9 -> 82.0 | No clear effect |
| add object location mask to relative network input (ablation) | 3D-RelNet · SUNCG (ground-truth boxes, feature ablation) | scale accuracy %: 56.1 -> 69.7 | Improved |
| add object location mask to relative network input (ablation) | 3D-RelNet · SUNCG (ground-truth boxes, feature ablation) | translation accuracy %: 82.7 -> 84.0 | Improved |
| add spatial coordinates to input (ablation) | 3D-RelNet · SUNCG (ground-truth boxes, feature ablation) | rotation accuracy (err <= 30 deg %): 82.4 -> 86.6 | Improved |
| add spatial coordinates to input (ablation) | 3D-RelNet · SUNCG (ground-truth boxes, feature ablation) | scale accuracy %: 70.4 -> 69.7 | Hurt |
| replace per-object prediction with pairwise relationship reasoning | Factored3D · SUNCG (ground-truth boxes) | translation mean error (m): 0.39 -> 0.33 | Improved |
| add CRF relationship model | Factored3D · SUNCG (ground-truth boxes) | translation accuracy %: 79.5 -> 79.7 | No clear effect |
| use 3D-RelNet instead of GCN | GCN · SUNCG (ground-truth boxes) | scale accuracy %: 69.1 -> 69.7 | No clear effect |
| use 3D-RelNet with detector boxes | Factored3D · SUNCG (detection setting) | mAP (all criteria): 21.72 -> 27.76 | Improved |
| use 3D-RelNet with detector boxes | Factored3D · NYUv2 (detection setting) | mAP (all criteria): 5.30 -> 8.49 | Improved |
Closest papers by meaning
- Learning 3D Object Shape and Layout Without 3D Supervision (CVPR 2022 · similarity 0.88)
- Edge-Centric Relational Reasoning for 3D Scene Graph Prediction (AAAI 2026 · similarity 0.88)
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · similarity 0.87)
- NOPE: Novel Object Pose Estimation from a Single Image (CVPR 2024 · similarity 0.86)
- Object-Centric Representation Learning for Enhanced 3D Semantic Scene Graph Prediction (NeurIPS 2025 · similarity 0.86)
- Open3DSG: Open-Vocabulary 3D Scene Graphs from Point Clouds with Queryable Objects and Open-Set Relationships (CVPR 2024 · similarity 0.86)
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.
