Key-Grid: Unsupervised 3D Keypoints Detection using Grid Heatmap Features
Authors: Chengkai Hou, Zhengrong Xue, Bingyang Zhou, Jinghan Ke, Lin Shao, Huazhe Xu
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct an extensive evaluation of Key-Grid on a list of benchmark datasets. Key-Grid achieves the state-of-the-art performance on the semantic consistency and position accuracy of keypoints. In this section, we compare the performance of Key-Grid over the existing SOTA approaches on both rigid-body and deformable object datasets. |
| Researcher Affiliation | Collaboration | Chengkai Hou1,3, Zhengrong Xue1,2, Bingyang Zhou4, Jinghan Ke5, Lin Shao6, Huazhe Xu1,2,7 1Shanghai Qizhi Institute 2Tsinghua University 3Peking University 4 The University of Hong Kong 5 University of Science and Technology of China 6 National Unversity of Singapore 7 Shanghai AI Lab |
| Pseudocode | No | The paper describes its method in detailed text and mathematical equations (e.g., sections 3.1, 3.2, 3.3) but does not include a formally labeled 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | Yes | We are committed to releasing the code. We release the code to facilitate the reproducibility of our work for the readers. The code is already provided in the supplementary materials. |
| Open Datasets | Yes | We use the Shape Net Core V2 and the Clothes Net datasets [2; 49] to evaluate the performance of Key-Grid. For the Shape Net Core V2 dataset [2]... For the Clothes Net dataset [49]... |
| Dataset Splits | No | The paper mentions using Shape Net Core V2 and Clothes Net datasets and describes training parameters like epochs and input point cloud size, but does not explicitly state the training/validation/test dataset splits or percentages. |
| Hardware Specification | Yes | Table 6 presents the time and memory consumption of Key-Grid and SM when inferring a batch comprising 32 samples on a single 1080 Ti GPU. |
| Software Dependencies | No | We use the Py Torch framework [7] to train our method. |
| Experiment Setup | Yes | The entire network is optimized on the Adam optimizer [10] with a learning rate of 0.1. For all the experiments, the batch size is set to 8. Our method is trained for 100 epochs, and for the first 20 epochs of training, we do not use a similarity loss function which measures the similarity between the reconstructed point clouds and the original point clouds. In the process of building the grid heatmap, the number of points M uniformly selected on each edge of the cubic space is 16. For the decoder, the hyperparameter Nneig for the number of neighboring features applied in aligning the feature from the former layer to the current layer is set to 3. In the loss function, αfar and αsim are both set to 1. |