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..
SegGraph: Leveraging Graphs of SAM Segments for Few-Shot 3D Part Segmentation
Authors: Yueyang Hu, Haiyong Jiang, Haoxuan Song, Jun Xiao, Hao Pan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on Part Net-E demonstrate that our method outperforms all competing baselines by at least 6.9% m Io U. |
| Researcher Affiliation | Academia | Yueyang Hu1, Haiyong Jiang1 , Haoxuan Song1, Jun Xiao1 , Hao Pan2 1School of Artificial Intelligence, University of Chinese Academy of Sciences 2School of Software, Tsinghua University EMAIL EMAIL EMAIL |
| Pseudocode | No | The paper describes the methodology using textual explanations and architectural diagrams (e.g., Fig. 2), but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at: https://github.com/YueyangHu2000/SegGraph. |
| Open Datasets | Yes | We evaluate the effectiveness of our method on two benchmark datasets: Part Net-Ensemble (Part Net E) [6] and Shape Net Part [49]. |
| Dataset Splits | Yes | The dataset splits follow that of [12]. We evaluate our method on the official testing set using a few-shot setting, where 8-shot examples are sampled from the training set for all experiments. |
| Hardware Specification | Yes | All measurements are on a per-shape basis, with training time referring to one epoch for a single shape, using a single NVIDIA V100 GPU. |
| Software Dependencies | No | The paper mentions various foundation models and libraries used (e.g., DINOv2, SAM, GLIP) but does not explicitly state their version numbers or other specific software dependencies required for reproduction in the main text. |
| Experiment Setup | Yes | We evaluate our method on the official testing set using a few-shot setting, where 8-shot examples are sampled from the training set for all experiments. The image feature maps are upsampled to the original image resolution using bicubic interpolation and are mapped from 768 channels to 96 channels with a linear layer. We feed the segment feature F s to a three GATv2 network layer. For the network training, we use cross-entropy as the objective to train the network with few-shot examples. |