patchDPCC: A Patchwise Deep Compression Framework for Dynamic Point Clouds
Authors: Zirui Pan, Mengbai Xiao, Xu Han, Dongxiao Yu, Guanghui Zhang, Yao Liu
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The evaluation on the MPEG 8i dataset shows that our method improves the compression ratio by 47.01% and 85.22% when compared to PCGCv2 and V-PCC with the same reconstruction quality, which is 9% and 16% better than that D-DPCC does. Our method also achieves the fastest decoding speed among the learning-based compression models. |
| Researcher Affiliation | Academia | Zirui Pan1, Mengbai Xiao1*, Xu Han1, Dongxiao Yu1, Guanghui Zhang1, Yao Liu2 1Shandong University 2Rutgers University |
| Pseudocode | No | The complete algorithm of copying points to the Ppatch is shown in Appendix3. |
| Open Source Code | Yes | Our Appendix is at https://github.com/pzrsdu/patch DPCC |
| Open Datasets | Yes | We train patch DPCC, PCGCv2, Oct Attention with the Owlii dataset (Xu, Lu, and Wen 2017), which contains 4 point cloud sequences. [...] We test the learning-based and handcrafted rule-based models both with the MPEG 8i dataset (d Eon et al. 2017), which includes 4 point cloud sequences with the precision of 10-bit. |
| Dataset Splits | No | The paper uses the Owlii dataset for training and the MPEG 8i dataset for testing, but does not specify training/validation/test splits within either dataset or how cross-validation is performed. |
| Hardware Specification | Yes | The training and testing of all methods are carried out on a single server equipped with an NVIDIA RTX3090 GPU with 24GB GDDR memory. |
| Software Dependencies | No | The paper mentions using Adam optimizer but does not specify versions for programming languages or other software libraries. |
| Experiment Setup | Yes | To obtain different tradeoffs between the compression ratio and the reconstruction quality, we set λ in the loss function as one of {1, 1.5, 2, 2.5, 3} and train multiple compression models. The training parameter α linearly increases from 0.001 to 1, so the reconstruction focuses more on details at later epochs. In the training stage, the batch size is 5, and we use Adam optimizer (Kingma and Ba 2015) with a learning rate decaying from 10 3 to 10 6. We set K to 16 in neighbor searching, N to 2048 as the patch size, and G to 256 as the feature dimension. |