Non-rigid Point Cloud Registration with Neural Deformation Pyramid
Authors: YANG LI, Tatsuya Harada
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments 4.1 Benchmarking partial-to-partial non-rigid point cloud registration 4DMatch/4DLo Match benchamrk. 4DMatch/4DLo Match [19] is a benchmark for non-rigid point cloud registration. Table 1: Quantitative non-rigid registration results on 4DMatch and 4DLo Match. 4.2 Ablation study Pyramid level. We test 9 pyramid levels with the initial frequency parameters in Eqn. 2 set to k0 = 8. Fig. 5 shows that, 1) the first level only slightly surpasses the rigid registration baseline, 2) the registration accuracies gradually increase with the pyramid level, and 3) lower levels tend to need more iterations to converge. |
| Researcher Affiliation | Academia | Yang Li1 liyang@mi.t.u-tokyo.ac.jp Tatsuya Harada1,2 harada@mi.t.u-tokyo.ac.jp 1The University of Tokyo 2RIKEN |
| Pseudocode | Yes | Algorithm 1 Non-rigid registration using NDP |
| Open Source Code | Yes | Code is available at https://github.com/rabbityl/Deformation Pyramid. |
| Open Datasets | Yes | 4DMatch/4DLo Match [19] is a benchmark for non-rigid point cloud registration. It is constructed using animation sequences from Deforming Things4D [20]. |
| Dataset Splits | No | The paper states 'All supervised models are re-trained on 4DMatch s training split before evaluation.' but does not explicitly provide details of a validation split used for their method. |
| Hardware Specification | Yes | with GPU accerleration (NVIDIA A100) |
| Software Dependencies | No | The paper mentions 'Point-to-point Iterative Closest Point (ICP) [2] implemented in Open3D [39]' and other libraries implicitly but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | Optimization of an MLP stops if 1) the max_iter = 500 is reached, 2) a given registration cost threshold γ = 0.0001 is reached, or 3) the registration cost does not change for more than σ = 15 iterations. We test 9 pyramid levels with the initial frequency parameters in Eqn. 2 set to k0 = 8. for each pyramid level we use a small MLP of (witdth, depth) = (128, 3). The registration parameters (m, k0) = (9, 8). |