دانلود Few-Shot Deep Adversarial Learning for Video-Based Person Re-Identification

Abstract

Video-based  person  re-identification  (re-ID)  refers to matching people across camera views from arbitrary unaligned video   footages. Existing  methods   rely on   supervision   signals to  optimise   a projected space   under   which   the distances between  inter/intra-videos  are  maximised/minimised.  However, this demands exhaustively labelling people across camera views, rendering  them  unable  to  be  scaled  in  large  networked cam-eras. Also, it is noticed that learning effective video   representations  with  view  invariance  is  not  explicitly  addressed  for which  features  exhibit  different distributions  otherwise.  Thus, matching  videos  for  person  re-ID  demands  flexible  models  to capture   the  dynamics   in time-series  observations   and   learn view-invariant   representations   with   access   to   limited   labeled training  samples.  In  this  paper,  we  propose  a  novel  few-shot deep  learning  approach  to  video-based  person  re-ID,  to  learn comparable  representations  that  are  discriminative  and  view-invariant.  The  proposed  method  is  developed  on  the  variational recurrent  neural  networks  (VRNNs)  and  trained  adversarially to produce latent variables with temporal dependencies that are highly  discriminative  yet  view-invariant  in  matching  persons. Through  extensive  experiments  conducted  on  three  benchmark datasets,  we  empirically  show  the  capability  of  our  method  in creating  view-invariant  temporal  features  and  state-of-the-art performance  achieved  by our  method.