Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Sparis: Neural Implicit Surface Reconstruction of Indoor Scenes from Sparse Views
Authors: Yulun Wu, Han Huang, Wenyuan Zhang, Chao Deng, Ge Gao, Ming Gu, Yu-Shen Liu
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experiments conducted on widely used benchmarks demonstrate superior performance in sparse-view scene reconstruction. Our extensive evaluations on both real-world and synthetic datasets show that Sparis achieves superior performance over current leading indoor reconstruction methods with sparse views. |
| Researcher Affiliation | Academia | 1Beijing National Research Center for Information Science and Technology (BNRist), Tsinghua University, Beijing, China 2School of Software, Tsinghua University, Beijing, China EMAIL, EMAIL |
| Pseudocode | No | The paper describes methods and processes through text and diagrams (Figure 2), but does not contain explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing source code or provide a link to a code repository for the methodology described. |
| Open Datasets | Yes | Scan Net. Scan Net (Dai et al. 2017), a comprehensive realworld dataset, encompasses over 2.5 million views across 1513 scenes... Replica. The Replica dataset (Straub et al. 2019) is notable for its high-quality reconstructions of various indoor environments. |
| Dataset Splits | No | Scan Net. ...sampling 15 to 20 images per scene at a resolution of 624 468 for surface reconstruction. Replica. ...10 images are uniformly sampled out of 2000, at a resolution of 600 340 for our experimental dataset. The paper specifies the number of input images used per scene but does not provide explicit training, validation, or test dataset splits for model training. |
| Hardware Specification | Yes | All the experiments are conducted on an NVIDIA RTX3090 GPU. |
| Software Dependencies | No | We adopt a similar model architecture as Vol SDF (Yariv et al. 2021). Ro Ma (Edstedt et al. 2023), a robust network for dense matching, is adopted as network fϕ to compute priors between images. We utilize the pre-trained Omnidata (Eftekhar et al. 2021) as our normal estimation network fθ to generate monocular normal priors. The paper mentions several software components and networks but does not provide specific version numbers for any of them. |
| Experiment Setup | No | Mono SDF, under its default hyper-parameter configuration, was unable to produce valid meshes; thus, we modified the weight of the monocular depth loss to 0.001 (originally 0.1) for a more equitable comparison. More experimental settings and metrics calculations are provided in the supplementary materials. The paper refers to a modification of a baseline's hyperparameter and states that more experimental settings are in the supplementary materials, but does not provide specific hyperparameters or system-level training settings in the main text. |