CAT-3DGS: A Context-Adaptive Triplane Approach to Rate-Distortion-Optimized 3DGS Compression
Authors: Yu-Ting Zhan, Cheng-Yuan Ho, He-Bi Yang, Yi-Hsin Chen, Jui Chiu Chiang, Yu-Lun Liu, Wen-Hsiao Peng
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 EXPERIMENTAL RESULTS: We conduct ablation experiments on the Mip-Ne RF360 dataset for its diverse scenes. |
| Researcher Affiliation | Academia | 1National Yang Ming Chiao Tung University, Taiwan 2National Chung Cheng University, Taiwan |
| Pseudocode | No | The paper describes methods and processes verbally without presenting them in structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not explicitly state that source code for the described methodology is available, nor does it provide a link to a code repository. |
| Open Datasets | Yes | We follow the common test protocol to test our CAT-3DGS on real-world scenes, including Mip-Ne RF 360 (Barron et al., 2022a), Tanks & Temples (Knapitsch et al., 2017), Deep Blending (Hedman et al., 2018) and Bungee Ne RF (Xiangli et al., 2022). |
| Dataset Splits | No | We follow the common test protocol to test our CAT-3DGS on real-world scenes, including Mip-Ne RF 360 (Barron et al., 2022a), Tanks & Temples (Knapitsch et al., 2017), Deep Blending (Hedman et al., 2018) and Bungee Ne RF (Xiangli et al., 2022). For comparison, we choose the same scenes from each dataset as those used in the prior works (Lu et al., 2024; Chen et al., 2024). |
| Hardware Specification | Yes | This information is collected on one NVIDIA V100. |
| Software Dependencies | No | The paper mentions various models and frameworks (e.g., Scaffold GS, 3DGS, HAC, Context GS) but does not provide specific version numbers for any software dependencies or programming languages used. |
| Experiment Setup | Yes | First, the spatial resolution B of the triplane at the lowest scale (r = 1) is determined in proportional to the number of anchor points obtained after 10k training iterations. The choices of the other hyperparameters include: the channel number ch = 72, ϵ = 0.01 (0.0004 for Bungee Ne RF) for the view frequency-aware masking, M = 4 with uneven slices (5, 10, 15, 25) for the channel-wise autoregressive coding. The rate parameter λr ranges from 0.002 to 0.04, and from 0.001 to 0.02 for Bungee Ne RF. Lastly, our triplanes have only two scales; that is, r = 1, 2. |