A Comparative analysis for Pair Re-ID

Code repository relating to master thesis "A Comparative analysis of re-ID models for matching pairs of Identities" by Nathan Bastien from EPL 2020 MAP promotion (inchallah).

Problem definition and objective of the models

• Pair Re-ID question: Is this pair of images, each containing a person, from the same identity ?
• Objective of the models: The model learns a feature space where pairs of embeddings of the same identity are separated by a smaller distance than pairs of embeddings of different identities.

Models

Trains 5 models for Pair Re-ID:

• CL: classification supervision
• ML: metric learning supervision
• CL2ML: adds a metric learning supervised model on top of the CL CNN. 2 versions of this models:
  • Multi-layer perceptron
  • Mahalanobis distance
• CL+ML model: supervision with classification and metric learning objectives.

Requirement

• Libraries see env.txt file
• Market-1501 dataset (http://www.liangzheng.com.cn/Project/project_reid.html)
• Keep file structure:
  -repo
  --- output
  --- dataset

Background

• Master thesis paper (in writing)
• Builds mainly from following papers:
  • Rethinking Person Re-Identification with Confidence (https://arxiv.org/abs/1906.04692)
  • Bag of Tricks and A Strong Baseline for Deep Person Re-identification (https://arxiv.org/abs/1903.07071)
  • In Defense of the Triplet Loss for Person Re-Identification (https://arxiv.org/abs/1703.07737)
  • Online Deep Metric Learning (https://arxiv.org/abs/1805.05510)

Code Structure

1. config.py: Defines all directories names and paths to outputs of the program
2. build_dataset.py: Splits Market-1501 in training, validation, query and gallery sets
3. train.py: trains the CNNS using train_utils.py and loss.py
4. extract_features.py: Extract features from dataset (training or testing) If CL2ML model:
   4.b) go to 5) OR
   4.b) train_mlp.py: Train MLP from training set extracted features Metric Learning OR
   4.b) train_moml.py: Train Matrix from extracted features for Mahalanobis distance based Metric Learning
   4.c) compare_metrics.py: Compare Metric learning training performances on training set
5. evaluate.py: Compute distances between all possible pairs of one query image and one gallery image. Also outputs traditional Re-ID metrics
6. dist_distribution.py: Sorts distances between same ID pairs and different IDs pairs. Output pairs examples from tails of distribution