Shape Context for soft biometrics in person re-identification and database retrieval

• Two scenarios of application for surveillance applications : re-identification and retrieval.
• Novel Shape Context (SC) descriptor, obtained from the head-torso silhouettes.
• The SC descriptor efficiently performs a robust soft biometric feature (BF) estimation via regression.
• Extensive experimental evaluation concerning both synthetic data (avatars) and real data.
• Remarkable performance is achieved in both scenarios using the concept of SC-BF.

We introduce a novel descriptor for the analysis of pedestrians and its applications to person re-identification and database retrieval. A Shape Context descriptor of the head-torso region of persons’ silhouettes is shown to have a very good discrimination ability and application to re-identification. For database retrieval using human queries, we train a map from the Shape Context to interpretable soft biometric quantities that can be reasoned about by humans. We show that a good linear correlation exists between Shape Context descriptors and soft biometrics quantities in the upper human torso and illustrate its application to retrieval in databases from human queries. Shape Context to biometrics maps are learned from virtual avatars rendered by computer graphics engines, to circumvent the need for time-consuming manual labelling of data sets. We obtained promising results of Shape Context based person re-identification and database retrieval from human compliant description of biometric traits, in both synthetic data and real imagery.