Learning Consistency-Aware Unsigned Distance Functions Progressively from Raw Point Clouds
Authors: Junsheng Zhou, Baorui Ma, Yu-Shen Liu, Yi Fang, Zhizhong Han
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experimental results in surface reconstruction for synthetic and real scan data show significant improvements over the state-of-the-art under the widely used benchmarks. Project page: https://junshengzhou.github.io/CAP-UDF. ... 4 Experiments We evaluate our method on the task of surface reconstruction from raw point clouds. We first demonstrate the ability of our method to reconstruct general shapes with open and multi-layer surfaces in Sec.4.1. Next, we apply our method to reconstruct surfaces for real scanned raw data including 3D objects in Sec.4.2 and complex scenes in Sec.4.3. Ablation studies are shown in Sec.4.4. |
| Researcher Affiliation | Academia | Junsheng Zhou1 Baorui Ma1 Yu-Shen Liu1 Yi Fang2 Zhizhong Han3 School of Software, BNRist, Tsinghua University, Beijing, China1 Center for Artificial Intelligence and Robotics, New York University Abu Dhabi, Abu Dhabi, UAE2 Department of Computer Science, Wayne State University, Detroit, USA3 |
| Pseudocode | No | The paper describes algorithms and methods in prose and mathematical formulations but does not include any explicitly labeled pseudocode blocks or algorithm listings. |
| Open Source Code | Yes | Project page: https://junshengzhou.github.io/CAP-UDF. ... Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] |
| Open Datasets | Yes | Dataset and metrics. For the experiments on synthetic shapes, we follow NDF [12] to choose the Car" category of the Shape Net dataset which contains the greatest amount of multi-layer shapes and non-closed shapes. ... Besides, we employ the MGD dataset [4] to show the advantage of our method in open surfaces. ... For surface reconstruction of real point cloud scans, we follow SAP to evaluate our methods under the Surface Reconstruction Benchmarks (SRB) [56]. ... To further demonstrate the advantage of our method in surface reconstruction of real scene scans, we follow On Surf [37] to conduct experiments under the 3D Scene dataset [63]. |
| Dataset Splits | No | The paper describes its training strategy for individual point clouds and uses various datasets for evaluation, but it does not specify explicit training, validation, or test dataset splits (e.g., percentages or sample counts) for the overall datasets used in experiments. It states: "given the single point cloud P as input, we do not leverage any condition and overfit the network to approximate the surface of P by minimizing the loss of Eq. (3). Therefore, we do not need to train our network on large scale training dataset". |
| Hardware Specification | No | The paper indicates that experiments were run and reports results but does not specify any particular hardware details such as GPU models, CPU types, or memory configurations used for the experiments. |
| Software Dependencies | No | The paper mentions using a "neural network similar to Occ Net [40]" but does not specify any software dependencies or libraries with their version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | Implementation details. To learn UDFs for raw point clouds P, we adopt a neural network similar to Occ Net [40] to predict the unsigned distance given 3D queries as input. Our network contains 8 layers of MLP where each layer has 256 nodes. Similar to Neural-Pull and SAL, given the single point cloud P as input, we do not leverage any condition and overfit the network to approximate the surface of P by minimizing the loss of Eq. (3). ... In addition, we use the same strategy as Neural-Pull to sample 60 queries around each point pi on P as training data. A Gaussian function N(µ, σ2) is adopt to calculate the sampling probability where µ = pi and σ is the distance between pi and its 50-th nearest points on P. For sampling auxiliary points in the low confidence region, the standard deviation is set to 1.1σ. And we train our network for two stages in practice. |