SCP: Spherical-Coordinate-Based Learned Point Cloud Compression
Authors: Ao Luo, Linxin Song, Keisuke Nonaka, Kyohei Unno, Heming Sun, Masayuki Goto, Jiro Katto
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that SCP surpasses previous state-of-the-art methods by up to 29.14% in point-to-point PSNR BD-Rate. The paper includes dedicated sections like 'Experiments', 'Experiment Settings', 'Experimental Results', and 'Ablation Study' where it evaluates models on datasets, reports metrics, and compares against baselines. |
| Researcher Affiliation | Collaboration | Ao Luo1,2, Linxin Song2, Keisuke Nonaka1, Kyohei Unno1, Heming Sun3, Masayuki Goto2, Jiro Katto2 1KDDI Research, Inc. 2Waseda University 3Yokohama National University |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions reproducing baseline methods (e.g., 'our reproduced EHEM... as the official implementation of EHEM is not provided') but does not provide an explicit statement or link to the source code for the SCP method itself. |
| Open Datasets | Yes | We compare all the baseline models and SCP on two Li DAR PC datasets, Semantic KITTI (Behley et al. 2019) and Ford (Pandey, Mc Bride, and Eustice 2011). |
| Dataset Splits | No | The paper specifies training and evaluation splits (e.g., 'The default split for training includes sequences 00 to 10, while sequences 11 to 21 are used for evaluation' for Semantic KITTI), but it does not explicitly define a separate 'validation' split distinct from the 'test' or 'evaluation' set in the dataset partitioning details. |
| Hardware Specification | Yes | All the experiments are done on a server with 8 NVIDIA A100 GPUs and 2 AMD EPYC 7742 64-core CPUs. |
| Software Dependencies | No | The paper mentions using the 'default Adam optimizer' and indicates that experiments are conducted using 'Oct Attention' and 'EHEM' models, but it does not provide specific version numbers for these or other software dependencies like Python, PyTorch/TensorFlow, or CUDA. |
| Experiment Setup | Yes | We train SCP models on Semantic KITTI and Ford datasets for 20 and 150 epochs, respectively. We employ the default Adam optimizer (Kingma and Ba 2014) for all experiments with a learning rate of 10-4. During training, we input the information of Spherical-coordinate-based Octree to the models, including the following information: (1) Octant: the integer order of the current voxel in the parent voxel, in the range of [1, 8]; (2) Level: the integer depth of the current voxel in the Octree, in the range of [1, 16] and [1, 17] for Semantic KITTI and Ford, respectively; (3) Occupancy, Octant, Level of ancient voxels: All the information of former levels. We trace back 3 levels... (4) Position: three floating-point positions of the current voxel in the Spherical coordinates, regularized to range [0, 1]. |