You look so different! Haven’t I seen you a long time ago?

Abstract

Person re-identification (re-id) aims to match a query identity (ID) to an element in a gallery set, composed of elements collected from multiple cameras. Most of the existing re-id methods assume the short-term setting, where the query/gallery samples share the clothing style. In this setting, the optimal feature representations are based in the visual appearance of clothes, which considerably drops the identification performance for long-term settings. Having this problem in mind, we propose a model that learns long-term representations of persons by ignoring any features previously learned by a short-term re-id model, which naturally makes it invariant to clothing styles. We start by synthesizing a data set in which we distort the most relevant biometric information (based in face, body shape, height, and weight cues), keeping the short-term cues (color and texture of clothes) unchanged. This way, while the original data contains both ID-related and other varying features, the synthesized representations are composed mostly of short-term attributes. Then, the key to obtaining stable long-term representations is to learn embeddings of the original data that maximize the dissimilarity with the previously inferred short-term embeddings. In practice, we use the synthetic data to learn a model that embeds the ID-unrelated features and then learn a second model from the original data, where long-term embeddings are obtained, keeping their independence with respect to the previously obtained ID-unrelated features. Our experiments were performed on three challenging cloth-changing sets (LTCC, PRCC, and NKUP) and the results support the effectiveness of the proposed method, for both short and long-term re-id settings. The source code is available at https://github.com/Ehsan-Yaghoubi/You-Look-So-Different-Haven-t-I-Seen-You-a-Long-Time-Ago?