Learning 3D Dense Correspondence via Canonical Point Autoencoder
Authors: An-Chieh Cheng, Xueting Li, Min Sun, Ming-Hsuan Yang, Sifei Liu
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on 3D semantic keypoint transfer and part segmentation transfer show that our model performs favorably against state-of-the-art correspondence learning methods. |
| Researcher Affiliation | Collaboration | An-Chieh Cheng1, Xueting Li2, Min Sun13, Ming-Hsuan Yang245, Sifei Liu6, 1National Tsing-Hua University 2University of California, Merced, 3Joint Research Center for AI Technology and All Vista Healthcare, 4Google Research, 5Yonsei University, 6NVIDIA |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code and trained models can be found at https://anjiecheng.github.io/cpae/. |
| Open Datasets | Yes | We carry out the semantic keypoint transfer task on the Keypoint Net dataset [46] as the BHCP benchmark used in Liu et al. [12] is not publicly available. Compared to the BHCP benchmark, the Keypoint Net dataset is more challenging because: (a) it contains diverse objects and comes with large-scale annotations, (b) it is template-free and annotated by a large group of people, thus is less biased compared to the keypoints in the BHCP benchmark, which are from predefined templates. For the part segmentation label transfer task, we use the Shape Net part dataset [47] as in [12]." and references [46] and [47] in bibliography. |
| Dataset Splits | Yes | In all experiments, including our method and the baselines, we use a validation set for model selection." and "For both datasets, we use the split provided in the original paper, and generate all pairs of shapes in the testing set as our testing pairs. |
| Hardware Specification | No | The paper states 'We thank National Center for High-performance Computing (NCHC) for providing computational and storage resources.' but does not provide specific hardware details like GPU/CPU models or memory. |
| Software Dependencies | No | The paper mentions 'the parameters of the network are optimized using the Adam [48] optimizer' but does not provide specific version numbers for any software dependencies like deep learning frameworks, libraries, or programming languages. |
| Experiment Setup | Yes | The training phase of our approach consists of two stages: (1) A pre-training stage trained with LACD (Eq. 1) and Lrec (Eq. 2) using α = 1 for LACD (2) A fine-tuning stage trained with LACD, Lrec, and Lcross (Eq. 3) where we set α = 0 for LACD. The total loss is formulated as Ltotal = ω1LACD +ω2Lrec+ω3Lcross, where ω1 = 10 for the pre-training stage; 20 for fine-tuning stage, ω2 = 1, ω3 = 10. For all experiments, we set k = 2048, τ = 0.9 (see Section 3.2), and the parameters of the network are optimized using the Adam [48] optimizer, with a constant learning rate of 1e 4. Also, for Lrec(P, S), µ1 = 1e3, µ2 = 10, and µ3 = 1. |