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..
Open-Vocabulary Part Segmentation via Progressive and Boundary-Aware Strategy
Authors: Xinlong Li, Di Lin, Shaoyiyi Gao, Jiaxin Li, Ruonan Liu, Qing Guo
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on Pascal-Part-116, ADE20K-Part-234, Part Image Net demonstrate that PBAPS significantly outperforms state-of-the-art methods, achieving 46.35% mIoU and 34.46% bIoU on Pascal-Part-116. Our method achieves state-of-the-art mIoU and bIoU scores on all three datasets, surpassing OVDiff and RIM. Table 3: Ablation study for PBAPS on Pascal-Part-116. |
| Researcher Affiliation | Academia | 1Tianjin University, China 2Southwest University, China 3Shanghai Jiao Tong University, China 4VCIP, CS, Nankai University, China |
| Pseudocode | No | The paper describes its methodology in prose and mathematical formulations within Section 3 "Boundary Feature Gradient Analysis" and Section 4 "Progressive Boundary-Aware Part Segmentation" but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/TJU-IDVLab/PBAPS. |
| Open Datasets | Yes | We evaluate PBAPS on three benchmarks: (1) Pascal-Part-116 [21], which refines the Pascal-Part [53] by merging over-segmented parts and removing redundant descriptors. (2) ADE20K-Part-234 [21], derived from ADE20K [54] via low-frequency class filtering and synonym merging, containing 44 object classes, 234 part classes, and 1016 validation images. (3) Part Image Net [33], which groups 158 Image Net [55] classes into 11 superclasses with uniform part structures, follows prior work [28] to evaluate 40 common object categories on 2957 validation images. |
| Dataset Splits | Yes | The validation set includes 17 object classes, 116 part classes, and 850 images. (2) ADE20K-Part-234 [21] [...] containing 44 object classes, 234 part classes, and 1016 validation images. (3) Part Image Net [33] [...] to evaluate 40 common object categories on 2957 validation images. |
| Hardware Specification | Yes | All results are measured with a single NVIDIA A6000 GPU. |
| Software Dependencies | Yes | Stable Diffusion v1.4 [30]... Vi T-B SAM [31]... Vi T-B DINOv2 [32]. |
| Experiment Setup | Yes | The hyperparameters of the BAR module are fixed as follows: ambiguity threshold λamb = 0.3, feature fusion weight γ = 0.8, and prototype adaptation coefficient α = 0.7. Finally, K-means clustering (K = 4) is applied to these features to construct subcategory prototypes. |