Search papers, labs, and topics across Lattice.
CoGe-GCD reframes Generalized Category Discovery (GCD) as a compositional generalization problem by inserting a modular inductive-bias layer between visual backbones and projection heads that maps patch tokens to a discrete primitive vocabulary. Rather than relying on monolithic, unstructured token embeddings that fail to extrapolate out-of-distribution, the framework pairs competitive token-primitive assignment with a structure-preserving spatial calibration to capture unseen primitive combinations. Across standard benchmarks, this plug-and-play approach consistently improves all-class accuracy and novel category number estimation with marginal computational overhead.
Novel classes are rarely entirely novel鈥攖reating them as unseen compositions of known visual primitives enables robust open-world category discovery without modifying downstream heads or objective functions.
Generalized Category Discovery (GCD) assigns unlabeled instances, mixed with labeled data, to known or novel categories, requiring human-like compositional reasoning: reusing primitives learned from known classes and deciding when new combinations imply new categories. Existing GCD methods operate on unstructured token features and struggle to extrapolate to novel compositions. We propose CoGe-GCD, which rethinks GCD through compositional generalization with two coupled stages. (i) Compositional Perception structures patch tokens by mapping them to a small vocabulary of primitives and refining token embeddings via competitive token-primitive assignment and information passing, yielding coherent groups for discovery. (ii) Generalizing Induction exploits the induced geometric structure and applies a structure-preserving calibration over spatial relations, maintaining probabilistic semantics while improving extrapolation to unseen primitive combinations. CoGe-GCD is implemented as an inductive-bias module between backbone and projection head, without modifying heads or losses, and can be plugged into diverse GCD frameworks. On standard benchmarks, it consistently improves all-class accuracy, unknown-class number estimation, and geometric quality, with marginal computational overhead. Code is available at https://github.com/lytang63/CoGe-GCD.
