Cross Domain Person Re-Identification With Large Scale Attribute Annotated Datasets

We propose a novel deep convolutional neural network framework called deep augmented attribute network (DAAN) to learn augmented attribute features for cross-domain person re-identification (person Re-ID) task. We observed that in some cases, different persons could have similar attributes (e.g., wearing similar clothes). It motivates us to distinguish such pedestrians by further learning complementary image features. We thus construct a deep neural network with three branches: (1) the attribute branch predicts the attributes of the input image; (2) the augmentation branch generates complementary features that are fused with the output of the attribute branch to form the augmented attribute features; and (3) the reconstruction branch to refine augmented attribute features on the target dataset. In order to learn precise and detailed attributes for pedestrian, we manually labeled two large datasets (CUHK03 and Market-1501) with 25 pre-defined mid-level semantic attributes. We evaluate the DAAN on a series of cross-domain person Re-ID tasks, where DAAN demonstrates superior performance (around 6%) to the prior state-of-the-art cross-domain algorithms.

Abstract licence: CC BY-NC-ND

40 more references, whose wording the paper’s licence keeps private.

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.