TL;DR: We present PartCo, a framework that uses part-level visual cues to enhance Generalized Category Discovery, achieving state-of-the-art performance with seamless integration into existing methods.
The focus of this paper is on Generalized Category Discovery (GCD). Given a labeled subset contains seen classes, the task is to categorize the unlabeled images, which may belong to seen or unseen classes.
Existing GCD methods primarily depend on semantic labels and global image representations, often overlooking the detailed part-level cues that are crucial for distinguishing closely related categories. In this paper, we introduce PartCo, short for Part-Level Correspondence Prior, a novel framework that enhances category discovery by incorporating part-level visual feature correspondences. By leveraging part-level relationships, PartCo captures finer-grained semantic structures, enabling a more nuanced understanding of category relationships. Importantly, PartCo seamlessly integrates with existing GCD methods without requiring significant modifications. Our extensive experiments on multiple benchmark datasets demonstrate that PartCo significantly improves the performance of current GCD approaches, achieving state-of-the-art results by bridging that are crucial for distinguishing closely related categories. In this paper, we introduce PartCo, short for Part-Level Correspondence Prior, a novel framework that enhances category discovery by incorporating part-level visual feature correspondences. By leveraging part-level relationships, PartCo captures finer-grained semantic structures, enabling a more nuanced understanding of category relationships. Importantly, PartCo seamlessly integrates with existing GCD methods without requiring significant modifications. Our extensive experiments on multiple benchmark datasets demonstrate that PartCo significantly improves the performance of current GCD approaches, achieving state-of-the-art results by bridging the gap between semantic labels and part-level visual compositions.
Object parts are often shared across categories and thus aid generalization to novel classes, yet many recent methods rely on transformer CLS tokens that capture global semantics while discarding crucial part-level details needed to distinguish similar categories. Although spatial prompt tuning (e.g., Wang et al., 2024) injects local cues, it struggles with inherent part variability in scale, pose, occlusion, and fluctuating part counts. This motivates an explicit, part-aware prior that is robust to these factors.
Overview of part-level correspondence labels construction: This two-step process begins by applying PCA projections to extract object and detailed features from ViT’s patch tokens using a subset of the dataset. These projections are then applied to the entire dataset to generate part-level correspondence labels.
(a) PartCo framework: Introduces part-level correspondence labels as a plug-and-play module to enhance GCD methods. (b) Part-level correspondence loss: This loss encourages features of the same part type and class to be close while separating different parts and/or classes, thereby enhancing the discriminative capability of the model at the part level.
We integrate our PartCo framework with the widely used parametric model: SimGCD (ICCV'23) and the SOTA non-parametric GCD method: SelEx (ECCV'24), employing DINO-variants backbones. Overall, PartCo integration consistently boost GCD methods, yielding new SOTA results across diverse datasets.
Attention map visualizations from PartCo show that our method consistently focuses on discriminative object parts, across both seen and novel classes, highlighting regions essential for category discovery.
This work is supported by the Hong Kong Research Grants Council - General Research Fund (Grant No.: 17211024).
@article{Cendra2025PartCo,
author = {Fernando Julio Cendra and Kai Han},
title = {PartCo: Part-Level Correspondence Priors Enhance Category Discovery},
journal = {arXiv preprint arXiv:2509.22769},
year = {2025}
}