Learning Deformable Hypothesis Sampling for Accurate PatchMatch Multi-View Stereo
Authors: Hongjie Li, Yao Guo, Xianwei Zheng, Hanjiang Xiong
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on DTU and Tanks & Temples datasets demonstrate its superior performance and generalization capabilities compared to state-of-the-art competitors. Ablation Studies are first conducted to independently validate the two modules of Deform Sampler (i.e., Plane Indicator Pθ and Probability Matcher Mθ). |
| Researcher Affiliation | Academia | The State Key Lab. LIESMARS, Wuhan University, China {lihongjie,guoyao gy,zhengxw,xionghanjiang}@whu.edu.cn |
| Pseudocode | Yes | The detailed DS-PMNet framework is presented in Algorithm 1 of the supplementary material. |
| Open Source Code | Yes | Code is available at https://github.com/Geo-Tell/DS-PMNet. |
| Open Datasets | Yes | Following the previous works like Trans MVSNet (Ding et al. 2022), we initially train our DSPMNet on the DTU dataset (Jensen et al. 2014) for DTU evaluation. Then, we fine-tune the model on the Blended MVS dataset (Yao et al. 2020), and test it on the Tanks and Temples benchmark (Knapitsch et al. 2017). |
| Dataset Splits | No | The paper mentions training on DTU and fine-tuning on Blended MVS, but does not explicitly describe specific validation dataset splits (e.g., percentages or counts for a validation set) for reproducibility. |
| Hardware Specification | Yes | The batch size is set to two on NVIDIA RTX 3090 for DTU and one for Blended MVS. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies (e.g., Python, PyTorch, CUDA versions) that would be needed to reproduce the experiment. |
| Experiment Setup | Yes | Our model is trained for 16 epochs with Adam optimizer, starting with a learning rate of 0.001, reduced by 0.2 at epochs 5, 9, and 13. To stabilize training against initial errors from random depth hypotheses, we set the initial learning rate of the probability matcher Mθ to 10 5. As for Fine-tuning on Blended MVS, our model undergoes 10 epochs with an initial learning rate of 0.0002, using 6 input images at a resolution of 576 768. The batch size is set to two on NVIDIA RTX 3090 for DTU and one for Blended MVS. |