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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Reasoning Beyond Points: A Visual Introspective Approach for Few-Shot 3D Segmentation
Authors: Changshuo Wang, Shuting He, Xiang Fang, Zhijian Hu, JIA-HONG HUANG, Yixian Shen, Prayag Tiwari
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on S3DIS and Scan Net demonstrate that our proposed VIP-Seg significantly outperforms current state-of-the-art methods, proving its effectiveness in PC-FSS tasks. Our code will be available at https://github.com/changshuowang/VIP-Seg . 4 Experiments 4.1 Datasets and Evaluation Metrics 4.2 Comparison with Existing Methods 4.3 Ablation Experiments |
| Researcher Affiliation | Academia | Changshuo Wang Department of Computer Science University College London London, United Kingdom EMAIL Shuting He School of Computing and Artificial Intelligence Shanghai University of Finance and Economics Shanghai, China EMAIL Xiang Fang Interdisciplinary Graduate Programme Nanyang Technological University, Singapore EMAIL Zhijian Hu LAAS CNRS Toulouse, France EMAIL Jia-Hong Huang Information Institute University of Amsterdam Amsterdam, Netherlands EMAIL Yixian Shen Information Institute University of Amsterdam Amsterdam, Netherlands EMAIL Prayag Tiwari School of Information Technology Halmstad University Halmstad, Sweden EMAIL |
| Pseudocode | No | The paper describes the components of the VIP-Seg network and its modules (Dy Power Conv, VIP Module) using descriptive text and mathematical formulations, and illustrates the overall architecture with block diagrams in Figure 2. However, it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | Our code will be available at https://github.com/changshuowang/VIP-Seg . |
| Open Datasets | Yes | We evaluate VIP-Seg on two widely adopted 3D segmentation benchmarks. S3DIS dataset [3] consists of RGB point clouds collected from 272 rooms across 6 indoor areas. ... Scan Net dataset [5] contains 1,513 scanned indoor scenes with dense point-wise annotations over 20 semantic categories... |
| Dataset Splits | Yes | Following the standard episodic learning paradigm [54], we partition the dataset into two nonoverlapping class sets: base classes Cbase used for training and novel classes Cnovel used for testing, where Cbase Cnovel = . The framework operates through N-way K-shot tasks, utilizing paired support and query sets. In each episode, the support set S = {(P n,k s , M n,k s )} contains K labeled samples for each of the N categories, where P n,k s RL (3+d) represents a point cloud with L points (each having 3D coordinates and d-dimensional features, such as color or surface normals), and M n,k s {0, 1}L denotes the corresponding binary segmentation mask indicating foreground (target class) and background points. The query set Q = {(Pq, Mq)} contains point clouds Pq RL (3+d) to be segmented, with ground truth masks Mq {0, 1, 2, ..., N}L assigning each point to one of the N target classes or the background class. |
| Hardware Specification | No | The paper discusses computational efficiency by comparing total training time (e.g., '0.9h of training') but does not specify any hardware details like GPU/CPU models, memory, or processor types used for the experiments. |
| Software Dependencies | No | The paper does not explicitly list any specific software dependencies with version numbers. |
| Experiment Setup | Yes | In Fig. 3(a), we analyze the effect of different reasoning steps in the VIP module. Performance improves from 1 step (68.74% m Io U) to 4 steps (72.20% m Io U), but decreases at 5 steps (71.12% m Io U), indicating optimal refinement occurs at 4 steps while excessive iterations cause over-processing. Fig. 3(b) shows the impact of Hi Conv layers. Performance increases from 1 layer (70.69% m Io U) to 3 layers (72.20% m Io U), then decreases with 4 layers (71.85% m Io U) and 5 layers (71.04% m Io U), suggesting 3 layers optimally balance feature capture and model complexity. Fig. 3(c) depicts the effect of encoder layers. The model improves dramatically from 1 layer (62.50% m Io U) to 3 layers (72.20% m Io U), gaining 9.70 percentage points, but declines with 4 layers (71.65% m Io U) and 5 layers (70.16% m Io U) due to overfitting. These results validate our architectural design choices. |