Research map / Point clouds / Registration and matching
DeepPRO: Deep Partial Point Cloud Registration of Objects
Donghoon Lee, Onur C. Hamsici, Steven Feng, Prachee Sharma, Thorsten Gernoth
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 point cloud generation network instead of naive global-feature regression (ablation) | DeepPRO · Linemod (unseen objects, avg of 11) | rotation error (deg): 2.12 -> 1.19 | Improved |
| use learned transform estimation network instead of SVD on dense correspondences (ablation) | DeepPRO · Linemod (unseen objects, avg of 11) | rotation error (deg): 1.94 -> 1.19 | Improved |
| add random rotation layer during training (ablation) | DeepPRO · Linemod (unseen objects, avg of 11) | rotation error (deg): 1.46 -> 1.19 | Improved |
| add bidirectional training path (ablation) | DeepPRO · Linemod (unseen objects, avg of 11) | rotation error (deg): 1.38 -> 1.19 | Improved |
| add ADD loss (ablation) | DeepPRO · Linemod (unseen objects, avg of 11) | rotation error (deg): 1.32 -> 1.19 | Improved |
| add ADD loss (ablation) | DeepPRO · Linemod (unseen objects, avg of 11) | ADD (cm): 0.27 -> 0.22 | Improved |
| replace DGCNN encoder with PointNet (ablation) | DeepPRO · Linemod (unseen objects, avg of 11) | rotation error (deg): 1.19 -> 1.36 | No clear effect |
| train with predicted object mask instead of ground truth mask (ablation) | DeepPRO · Linemod (unseen objects, avg of 11) | rotation error (deg): 1.19 -> 1.32 | No clear effect |
| reduce input points from 512 to 256 (ablation) | DeepPRO · Linemod (unseen objects, avg of 11) | rotation error (deg): 1.19 -> 1.26 | Hurt |
| reduce input points from 512 to 128 (ablation) | DeepPRO · Linemod (unseen objects, avg of 11) | ADD (cm): 0.22 -> 0.32 | Hurt |
| use DeepPRO instead of classical ICP | DeepPRO · Linemod (unseen objects, avg of 11) | rotation error (deg): 2.15 -> 1.19 | Improved |
| use DeepPRO instead of PRNet on synthetic ModelNet40 | PRNet · ModelNet40 (partial, noisy, large baseline) | RMSE(R): 4.323 -> 2.617 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- RPM-Net — RPM-Net: Robust Point Matching Using Learned Features (CVPR 2020) · better on 3 of 3 shared results
- PRNet — PRNet: Self-Supervised Learning for Partial-to-Partial Registration (NeurIPS 2019) · better on 2 of 2 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
- Deep Closest Point: Learning Representations for Point Cloud Registration (ICCV 2019 · similarity 0.91)
- DeepVCP: An End-to-End Deep Neural Network for Point Cloud Registration (ICCV 2019 · similarity 0.91)
- DeepI2P: Image-to-Point Cloud Registration via Deep Classification (CVPR 2021 · similarity 0.90)
- DeepMapping: Unsupervised Map Estimation From Multiple Point Clouds (CVPR 2019 · similarity 0.90)
- PCAM: Product of Cross-Attention Matrices for Rigid Registration of Point Clouds (ICCV 2021 · similarity 0.89)
- 3D Focusing-and-Matching Network for Multi-Instance Point Cloud Registration (NeurIPS 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.
