NeuralPlane: Structured 3D Reconstruction in Planar Primitives with Neural Fields
Authors: Hanqiao Ye, Yuzhou Liu, Yangdong Liu, Shuhan Shen
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments on Scan Netv2 and Scan Net++ demonstrate the superiority of our method in both geometry and semantics. |
| Researcher Affiliation | Academia | 1School of Artificial Intelligence, University of Chinese Academy of Sciences 2Institute of Automation, Chinese Academy of Sciences |
| Pseudocode | No | The paper describes methods in prose and mathematical equations but does not contain a clearly labeled 'Pseudocode' or 'Algorithm' block, nor structured steps formatted like code. |
| Open Source Code | Yes | Project page: https://neuralplane.github.io/ |
| Open Datasets | Yes | To evaluate Neural Plane s ability to reconstruct 3D plane maps, we conduct experiments on 12 challenging real-world indoor scenes: 8 scenes from Scan Netv2 (Dai et al., 2017) and 4 scenes from Scan Net++ (Yeshwanth et al., 2023). |
| Dataset Splits | Yes | we sample 8 scenes from the official validation set of Scan Netv2 and 4 scenes from Scan Net++ (Yeshwanth et al., 2023). ... For each scene from Scan Netv2, one-eighth of the frames are uniformly selected for reconstruction, while in Scan Net++, equally spaced frames are downsampled according to the number of registered frames, leaving 150 279 frames per scene. |
| Hardware Specification | Yes | Preprocessing local planar primitives takes around 2 to 5 minutes, followed by about 6 minutes for training on a single NVIDIA RTX 3090 GPU. |
| Software Dependencies | No | Neural Plane is implemented in Nerfstudio (Tancik et al., 2023) on top of Nerfacto, a unified approach in the literature of Ne RF. ... we use the COLMAP toolbox (Sarlin et al., 2019) that supports Lo FTR (Sun et al., 2021) , to export dense 3D keypoints for the initialization of local planar primitives. The paper mentions software by name but lacks specific version numbers for these or other libraries like PyTorch, Python, CUDA. |
| Experiment Setup | Yes | We train Neural Plane for 4k iterations with batches of 8192 rays across all scenes. ... We list several key hyperparameters used in our main experiments: the dimension of coplanarity feature d = 4, the pushing thresholds (to, tn) = (8 cm, cos 10 ), and the number of semantic prototypes Np = 32. ... During each training iteration, we randomly select 128 local planar primitives with probabilities weighted by the sizes of corresponding 2D plane segments. In each local planar primitive, we uniformly sample n = 8192 / 128 = 64 rays, with N=48 points sampled per ray. ... The push margin m in eq. (7) is set to 1.5 for the first 1k iterations, and is later increased to 2. In Neural Parser, the DBSCAN epsilon is fixed to 0.2. We optimize plane parameters of local planar primitives using Adam optimizer with an exponential decay schedule from an initial learning rate of 1 10 3 to 1 10 4. |