Differentiable Registration of Images and LiDAR Point Clouds with VoxelPoint-to-Pixel Matching
Authors: Junsheng Zhou, Baorui Ma, Wenyuan Zhang, Yi Fang, Yu-Shen Liu, Zhizhong Han
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experimental results on KITTI and nu Scenes datasets show significant improvements over the state-of-the-art methods. |
| Researcher Affiliation | Collaboration | Junsheng Zhou1 Baorui Ma1,2 Wenyuan Zhang1 Yi Fang3 Yu-Shen Liu1 Zhizhong Han4 School of Software, Tsinghua University, Beijing, China1 Beijing Academy of Artificial Intelligence2 New York University Abu Dhabi, Abu Dhabi, UAE3 Department of Computer Science, Wayne State University, Detroit, USA4 |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code and models are available at https://github.com/junshengzhou/VP2P-Match. |
| Open Datasets | Yes | We evaluate our performance on Image-to-Point Cloud registration task on two wildly used benchmarks KITTI and nu Scenes. KITTI Odometry [17]. We generate the image-point cloud pairs from the same data frame of 2D/3D sensors. We follow previous works [26] to utilize the 0-8 sequences for training, and 9-10 for testing. nu Scenes [4]. The image-point cloud pairs are generated by official SDK where the point cloud is accumulated from the nearby frames and the image is from the current frame. We follow the official data spilt of nu Scenes to utilize 850 scenes for training, and 150 scenes for testing. |
| Dataset Splits | No | The paper specifies train and test splits for KITTI and nu Scenes datasets, but does not explicitly mention a validation set split. |
| Hardware Specification | Yes | where all the methods are evaluated with an NVIDIA RTX 3090 GPU and Intel(R) Xeon(R) E5-2699 CPU. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., Python version, library versions like PyTorch, TensorFlow, or CUDA). |
| Experiment Setup | Yes | We set the channel dimension C of 2D/3D feature to 64, and set C2D and C3D for the 2D/3D global feature both to 512. We set the margin m to 0.25, the scale factor γ to 32 and the safe radius r to 1 pixel. To enhance the representation ability of the voxel branch, we keep the point transformation pipe used in SPVNAS [43]. And the probability threshold σ in intersection detection is set to 0.95. |