Person Re-Identification using EfficientNet and Triplet Loss

This project performs Person Re-Identification (Re-ID) using deep feature embeddings extracted via a pretrained EfficientNet-B0 model. The system retrieves visually similar identities from a gallery using cosine similarity on learned embeddings.

---

📁 Dataset

We used the Market-1501 dataset which consists of:

• 751 training identities
• 750 test identities
• Images captured from 6 camera views

---

🧠 Model Details

• Backbone: EfficientNet-B0
• Embedding Size: 512-dimensional vector
• Loss Function: Triplet Loss (using hard negative mining)
• Similarity Metric: Cosine similarity
• Training Epochs: 20+
• Framework: PyTorch

---

⚙️ How It Works

1. Each image is passed through a pretrained EfficientNet to extract a 512-d embedding.
2. A query image is compared to the embeddings of all gallery images using cosine similarity.
3. Top-K most similar images are returned as the result.
4. A Gradio web interface allows user interaction for live querying.

---