EMD-based Local Matching for Occluded Person Re-identification

This repository contains the implementation of EMD-based local matching for occluded Person Re-identification, as presented in our paper published in Machine Learning with Applications. The proposed method introduces a multiple-shot approach for Person Re-ID by integrating feature fusion and Earth Mover’s Distance (EMD)-based local matching.

📜 Paper

Title: EMD-based Local Matching for Occluded Person Re-identification
Authors: Hoang-Anh Nguyen, Thuy-Binh Nguyen, Hong-Quan Nguyen, Thi-Lan Le
Journal: Machine Learning with Applications

📄 Read the Paper (Add link when available)

🔥 Highlights

• Multi-shot Person Re-ID: Our framework extends single-shot Re-ID to a multiple-shot scenario by utilizing tracklet-based feature fusion.
• Tracklet Matching Approach: Person Re-ID is formulated as a pedestrian tracklet matching problem, where tracklets are generated via human detection and tracking.
• Occlusion Handling: The proposed framework tackles occlusions by incorporating EMD-based local matching.
• Ablation Study: The effectiveness of feature fusion is investigated to demonstrate its impact on Re-ID performance.

🚀 Methodology

The proposed method follows a three-step pipeline:

1. Feature Extraction

   • Each pedestrian tracklet is processed using ResNet-50 to extract global and local features.

2. Feature Fusion

   • Average pooling is applied to enhance person representation from multiple frames.

3. Matching using EMD

   • Global matching: Euclidean distance.
   • Local matching: Earth Mover’s Distance (EMD) for optimal alignment of local features.

The model is trained using identity loss and triplet loss for robust feature learning.

(Add a relevant image or diagram from the paper)

📊 Dataset

We conduct experiments on the FAPR dataset, which consists of pedestrian tracklets automatically extracted from videos captured by two non-overlapping cameras. Additionally, we evaluate the method on standard benchmarks like VIPeR and Market-1501-Partial.

📊 Results

Single shot

Table 1: Performance Comparison on VIPeR, Market-1501-Partial, and FAPR datasets

Method \ Matching rates	VIPeR (R-1)	VIPeR (R-5)	VIPeR (R-10)	Market-1501-Partial (R-1)	Market-1501-Partial (R-5)	Market-1501-Partial (R-10)	FAPR (R-1)	FAPR (R-5)	FAPR (R-10)
Global (DMLI)	37.8	68.2	79.9	68.6	83.2	87.6	58.0	88.0	92.0
Global (EMD)	38.3	67.8	79.7	74.7	87.2	91.2	60.2	89.4	95.1
Local (DMLI)	40.2	73.1	83.6	77.3	89.4	92.1	59.4	87.4	91.8
Local (EMD)	47.8	76.3	86.7	78.0	89.5	92.6	64.6	91.5	97.0
Global+Local (DMLI)	40.9	72.8	82.9	77.2	89.3	92.5	61.3	86.5	91.6
Global+Local (EMD)	47.8	76.6	86.5	78.0	89.1	92.6	65.8	89.4	95.8

Multi shot

Table 2: Performance Comparison on FAPR Dataset

Probe	Gallery	Baseline	Average
20191104 indoor_left	outdoor_left	55.3	100
20191104 indoor_right	outdoor_right	64.3	76.3
20191104 indoor_cross	outdoor_cross	45.5	100
20191104 indoor_all	outdoor_all	58.8	82.3
20191105 indoor_left	outdoor_left	100.0	100
20191105 indoor_right	outdoor_right	75.0	98.8
20191105 indoor_cross	outdoor_cross	57.1	98.9
20191105 indoor_all	outdoor_all	78.6	100

🔧 Installation

git clone https://github.com/your-repo/emd-person-reid.git
cd emd-person-reid
pip install -r requirements.txt

📂 Dataset Preparation

1. Download the FAPR dataset from link.
2. Extract the dataset and place it in the data/ directory.

data/
├── FAPR/
│   ├── query/
│   ├── gallery/

🏋️‍♂️ Training

Train the model using the following command:

python train_alignedreid_multi.py --dataset FAPR --epochs 100 --batch_size 32

📈 Evaluation

To evaluate the trained model:

python train_alignedreid_multi.py --dataset FAPR --checkpoint path/to/model.pth

🔗 Citation

If you use this work, please cite:

@article{nguyen2024emd,
  title={EMD-based Local Matching for Occluded Person Re-identification},
  author={Nguyen, Hoang-Anh and Nguyen, Thuy-Binh and Nguyen, Hong-Quan and Le, Thi-Lan},
  journal={Machine Learning with Applications},
  year={2024}
}

📬 Contact

For any questions or collaborations, please contact [[email protected]].