Towards Better Person Re-Identification through Interactive Visual Exploration and Incremental User Feedback

This repository contains source code used to explore and analyze the search space in person Re-Identification(Re-ID).

Introduction

We propose ReIDVis, which is a novel data exploration and data visualization prototype for person Re-ID.

ReIDVis integrates a user-feedback mechanism that incorporates the person Re-ID model with human insights, and a composite visualization that support efficient visual browsing, retrieval, and exploration of candidate targets. To help identify person-of-interest, we propose an extended semi-supervised learning method by introducing a k-fusion post-rank algorithm to support incremental user feedback. In the visualization component, we develop a novel cluster-based visualization with an optimized layout to reduce visual occlusion and preserve user’s mental map. We also propose a multi-scale, pixel-based view to guide user exploration in the search space.

System screenshots

Fig. 4: The interface of the system. (A) the probe panel which allows users to select the person-of-interest as a probe and set up the visual parameter of the search space. (B) the search space view which supports visual exploration and enables users to provide feedback on the samples. (C) the spatiotemporal information view which summarizes the spatiotemporal information of the retrieval results. (D) the ranking list composed of the pixel-based visual encoding which allows users to quickly retrieve strong negative samples. (E) the ranking list composed of the raw image. (F) the search space with the “route” mode, where a front photo of the probe(G) is found. (H) a node with pixel-based visual encoding.

The system demonstration video of ReIDVis addresses at：https://youtu.be/MjmH6cZYcj4

Function

• The semi-supervised learning method

  Input data: human-labeled data ('l_index' is the index of the labeled data in the ranking list, 'Yl' is the value of labels).

  label = {'l_index': [0, 2, 1, 3, 4], 'Yl': [1, 1, -1, -1, -1]}
  

  Output data: a new ranking list generated via LapSVM.py.

• The k-fusion post-rank algorithm

  Input data: branch ranking lists and a parameter k.

  Output data: a new ranking list generated by k-fusion.py.

• The cluster-based visualization

  • Ensure the stability of the incremental layout.

    Input data: the coordinates of samples at the previous time step t-1, the feature of samples at the current time step t.

    Output data: the coordinates of samples at the current time step t generated by ComputingWeight.py.

  • Remove the occlusion between nodes.

    Input data: the coordinate, width, and height of each node.

    Output data: the removed coordinate of each node generated by occlussion.py.

Installation

# cd to your preferred directory and clone this repo
git clone https://github.com/xiawang157/Video_Object_FeatureVis.git

# create dependencies
cd Video_Object_FeatureVis/
pip install -r requirements.txt

How to use

(1) Install and configure the development environment according to requirements.txt.

(2) Load data from , and put the data in static.

(3) Run views.py, and start the front-end web service.