[Paper note] Gated Siamese Convolutional Neural Network Architecture for Human Re-Identification

最新推荐文章于 2023-12-07 21:14:13 发布

原创最新推荐文章于 2023-12-07 21:14:13 发布 · 2.7k 阅读

CC 4.0 BY-SA版权

文章标签：

20 篇文章

订阅专栏

本文介绍了一种用于解决视角变化问题的人再识别(person re-id)模型。该模型通过匹配门结构(Matching Gate)来增强基线Siamese网络，并利用特征汇总技术提高不同视角下同一人的识别率。实验结果表明，此模型在Market-1501等多个数据集上取得了较好的性能。

Author: Rahul Rama Varior, NTU Singapore; Mrinal Haloi, Nanyang technological Universi; Gang Wang
Picutre of the model

2016_10_14_cebff7ce866e92a1c89e96dfc777d19

Feature summarization
- Aggregates the local feature along a horizontal stripe.
- Deal with problem of changed view point (view point change in re-id typically in the horizontal direction, same parts are very likely to be along the same horizontal region).
- Equation and dimention:
  $y_{r1}=f(w*x_{r1})$ ; $y_{r2}=f(w*x_{r2})$
  where $x_{ri}\in \mathbb{R}^{1\times c\times h}$ is the $r^{th}$ row of feature map. $w\in \mathbb{R}^{1\times c\times h\times h}$ (maybe $\mathbb{R}^{1\times c\times 1\times h}$ ) and $y_{r1}\in \mathbb{R}^{1\times 1\times h}$ (maybe $\mathbb{R}^{1\times 1\times h\times h}$ ).
Feature similarity
- Euclidean distance of $y_{r1}$ and $y_{r2}$
  $g_r^i=\exp(\frac{-(y_{r1}^i-y_{r2}^i)^2}{p_i^2})$
  where $g_r \in \mathbb{R}^{1\times 1\times h}$
- $p_i$ decides the variance of the Gaussian function, learnable, should set a higher initial value.
Filtering and boosting the features
- Repeat $g_r$ c times horizontally to obtain $G_r\in \mathbb{R}^{1\times c\times h}$ .
- Add filtered feature to original feature.
  $a_{r1}^i=x_{r1}^i+x_{r1}^i\odot G_r^i$
  $a_{r2}^i=x_{r2}^i+x_{r2}^i\odot G_r^i$
- Perform L2 normalization across channels after this

Dataset: Market-1501, CUHK-03, VIPeR.
Baseline S-CNN outperform most CNN approaches. With MG gaining further improvement.
Visualization of gate. Low gate activation means low similarity.