[3Rin4] Interactive 1

[3Rin4-74] Transfer Learning and its Visualization of Pose-invariant Domain-dependent Features in Person Re-identification

〇Yuki Murata¹, Masayasu Atsumi¹ (1.Soka University Graduate School of Engineering)

Keywords:Person Re-identification, Transfer Learning, Visualization

Person re-identification (re-id) is a problem to find the same person from person images captured by multiple cameras. We have proposed models that maintain the pose-invariant features obtained in the framework of FD-GAN and acquire domain-dependent features suitable for new data by transfer learning. In these model, additional re-id feature extraction encoders are introduced for the transfer learning of domain-dependent features. However, through experiments, we have found that it is difficult to achieve domain-dependent transferability while keeping the pose-invariant features. Accordingly, in this paper, we visualize the explanation of decision of the pose-invariant and domain-dependent re-identification by our proposed models using techniques such as CAM (Class Activation Mapping) for leading to get a clue of model architecture design.