Deep Sequential Multi-camera Feature Fusion for Person Re-identification
share
Given a target image as query, person re-identification systems retrieve a ranked list of candidate matches from other camera field-of-views. Re-identification is typically performed independently for each gallery camera across the network. However, images of the same target from a subset of cameras often provide complementary appearance information which, if aggregated judiciously, can lead to better re-id performance in the remaining cameras. This motivates us to investigate the novel problem of multi-camera feature fusion for person re-id. We propose an intelligent sequential fusion technique, designed to not only improve re-id accuracy but to also learn increasingly better feature representations as observations from additional cameras are fused. The fusion function is modeled using a Gated Recurrent Unit (GRU) with modification on default GRU training regime to achieve the aforementioned improvements. Extensive experimentation validates that the proposed fusion scheme substantially boosts identification performance when combined with off-the-shelf feature extraction methods for re-id.
READ FULL TEXT