Person Re-Identification Implementations

This repository contains two implementations for Person Re-Identification in video streams using machine learning techniques. The first is an older method based on image-based similarity comparison, while the second is a more recent approach leveraging YOLOv8 and Siamese networks for multi-frame analysis.

Implementations

1. Image-Based Re-Identification (Old Implementation)

In this implementation, we extract features from pairs of images and compare them to determine if they belong to the same person. The process involves:

• Pairwise comparison using a simple metric learning method (Mahalanobis Distance).

Results:
The results for this implementation are stored as two example images showcasing person re-identification:
Test Images


Similarity Ranks

2. YOLOv8 + Siamese Network (New Implementation)

This newer method integrates object detection with deep learning for more robust re-identification:

• YOLOv8 is used to detect and track people in video frames.
• A Siamese network is employed to compare feature vectors and compute similarity scores for re-identification.
• Multi-frame analysis and a dynamic reference management system handle occlusions and re-entries of individuals.

The results for this implementation are demonstrated in a video, which can be found at the following link: Watch the result video here

Requirements

To run the code for both implementations, install the following dependencies:

pip install -r requirements.txt

• PyTorch
• OpenCV
• YOLOv8

Running the Code

1. Old Implementation (Image-Based)

1. Install the required dependencies:

   pip install -r requirements.txt

2. Run the metric_learning.py script:

   python metric_learning.py

2. New Implementation (YOLOv8 + Siamese Network)

1. Install the required dependencies:

   pip install -r requirements.txt

2. Request the model weights by contacting the repository maintainer.
3. Once you have the weights, open and run inference.ipynb to perform person re-identification on video streams.

Conclusion

The new implementation using YOLOv8 and Siamese networks provides more robust and accurate results for person re-identification in video streams, particularly in dynamic and crowded environments. The video-based system also handles occlusions and re-entries, making it a significant improvement over the old image-based approach.

For detailed results, refer to the example images and the linked video.