Person re-identification (ReID) involves ranking an image gallery based on a query, which can be an image or text. In image-to-image (I2I) ReID, variations in viewpoint, illumination, occlusion, and background can cause distractors to appear more similar to the query than the true match.
The paper, titled "In PRAISE of Attributes," introduces a method that grounds retrieval geometry using query-conditioned subspaces. The approach appears to leverage attribute information to condition the feature space, aiming to make the retrieval more robust to the aforementioned variations.
By conditioning the subspace on the query, the method may better separate true matches from distractors. The abstract suggests that this grounding helps address the challenge of distractors appearing more similar under challenging conditions.