ElasticMVS: Learning elastic part representation for self-supervised multi-view stereopsis
Authors: Jinzhi Zhang, Ruofan Tang, Zheng Cao, Jing Xiao, Ruqi Huang, LU FANG
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive evaluations convey the superiority of Elastic MVS in the reconstruction completeness and accuracy, as well as the efficiency and scalability. Particularly, for the challenging large-scale reconstruction benchmark, Elastic MVS demonstrates significant performance gain over both the supervised and self-supervised approaches. |
| Researcher Affiliation | Collaboration | Jinzhi Zhang1,2 , Ruofan Tang1,3, Zheng Cao4, Jing Xiao5, Ruqi Huang2 and Lu Fang1 1Department of Electronic Engineering, Tsinghua University 2Tsinghua Shenzhen International Graduate School 3Dept. of Automation, Tsinghua University, 4Biren Tech Research, 5Pingan Group |
| Pseudocode | No | The paper describes the algorithms and procedures in prose (e.g., Section 3 'Method') but does not provide any structured pseudocode blocks or figures explicitly labeled 'Algorithm' or 'Pseudocode'. |
| Open Source Code | Yes | Code is avaliable at https://thu-luvision.github.io. |
| Open Datasets | Yes | DTU dataset [1] is an MVS dataset... Tanks and Temples [25] dataset contains two benchmarks... |
| Dataset Splits | No | The paper refers to using standard evaluation protocols for datasets like DTU and Tanks and Temples, but it does not explicitly provide specific training, validation, or test dataset splits (e.g., percentages or sample counts) within the paper itself. |
| Hardware Specification | Yes | All the training and reconstruction are conducted on a single NVIDIA GTX 3090 graphics card. |
| Software Dependencies | No | The paper mentions using a 'Feature Pyramid Network' and 'SGD' for training but does not provide specific version numbers for any software libraries, frameworks, or operating systems used in the experiments (e.g., 'PyTorch 1.9' or 'CUDA 11.1'). |
| Experiment Setup | Yes | The threshold ϵ for defining Sp is set to 0.5 in the DTU dataset and 0.01 in the T&T dataset. We fix the temperature number τ = 0.5 and the confidence threshold ξ = 0.7 during training and propagation. We sample Nc = 8 points for each pixel. During the propagation, we select Kp = 8 hypotheses for propagation and Kr = 4 for perturbation. We choose η = 0.1 in Eq. 2 to find nearby pixels. During the reconstruction stage, we set respectively αs = 0.1 and αg = 0.5 in Eq. 3 and 4. For the loss functions, β = 0.5 in Eq. 7 and γs = 0.005 in Eq. 8. |