3D-LMVIC: Learning-based Multi-View Image Compression with 3D Gaussian Geometric Priors
Authors: Yujun Huang, Bin Chen, Niu Lian, Xin Wang, Baoyi An, Tao Dai, Shu-Tao Xia
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that 3D-LMVIC achieves superior performance compared to both traditional and learning-based methods. Additionally, it significantly improves disparity estimation accuracy over existing two-view approaches. |
| Researcher Affiliation | Collaboration | 1Shenzhen International Graduate School, Tsinghua University, Shenzhen, China 2School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen, China 3Huawei Technologies Company Ltd., Shenzhen, China 4Department of Software Engineering, Shenzhen University, Shenzhen, China. |
| Pseudocode | Yes | Algorithm 1 Disparity and Mask Estimation Input: Depth estimation function GSDE, intrinsic matrix K, extrinsic matrices Vn and Vn 1 Output: Disparity map n, mask xn,m dn GSDE(K, Vn) dn 1 GSDE(K, Vn 1) n, d n 1 Disparity Estimation(dn, K, Vn, Vn 1) xn,m Mask Estimation( n, d n 1, dn 1) |
| Open Source Code | No | The paper does not contain an explicit statement about releasing source code for the described methodology, nor does it provide a link to a code repository. Mentions of software (HTM-16.3, HM-18.0) are for benchmark comparisons. |
| Open Datasets | Yes | We evaluate our model on three multi-view image datasets: Tanks&Temples (Knapitsch et al., 2017), Mip-Ne RF 360 (Barron et al., 2022), and Deep Blending (Hedman et al., 2018). |
| Dataset Splits | Yes | For all datasets, 90% of the images in each scene were allocated for training, with the remaining 10% used for testing. |
| Hardware Specification | Yes | These evaluations were conducted on a platform with an Intel(R) Xeon(R) Gold 6330 CPU @ 2.00GHz and a GPU containing 10,752 parallel processing cores. |
| Software Dependencies | No | The paper mentions general software components like 'Adam optimizer' and implicitly deep learning frameworks but does not specify version numbers for any libraries or specific software dependencies used in their method. |
| Experiment Setup | Yes | The model was trained using five different configurations of (λimg, λdep)... The weights wi for four consecutive views were set to (0.5, 1.2, 0.5, 0.9)... The model was trained for 300 epochs with an initial learning rate of 10 4, which was progressively decayed by a factor of 0.5 every 60 epochs. We utilize the Adam optimizer for training with a batch size of 2. |