Instance-Level Object Detection

This repository implements an instance-level object detection system developed as part of the TTIC 31040 Computer Vision course project.

🔍 Project Description

Given a reference image of an object and a test image, this system identifies all instances of the object in the test image. The reference image contains only the object of interest, and the object is not occluded.

The algorithm uses SIFT (Scale-Invariant Feature Transform) keypoints to detect and match features between the reference and test images. It then estimates geometric transformations to locate multiple instances in the scene.

Outputs

• ✅ The test image annotated with bounding boxes around detected instances.
• 📍 The (x, y) coordinates of the bounding box corners.
• 🔢 The total number of detected instances.

📁 Project Structure

• visualize.ipynb: Main notebook that runs the instance detection pipeline.
• helper_functions.py: Core implementation of feature detection and matching.
• requirements.txt: Python dependencies for the project.
• Sounak-Paul-FinalReport.pdf: Full project report with explanation and results.

⚙️ Setup Instructions

1. Clone the repository

git clone https://github.com/paulsounak96/instance-level-object-detect.git
cd instance-level-object-detect

2. Create and activate a virtual environment

python -m venv venv
# On Windows:
venv\Scripts\activate
# On macOS/Linux:
source venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. (Optional) Install Jupyter

pip install jupyter

▶️ Usage Instructions

1. Launch Jupyter Notebook:

jupyter notebook

2. Open visualize.ipynb and run all cells.

3. The notebook will:

   • Load the reference and test images.
   • Use SIFT to find matches.
   • Detect all instances of the object in the test image.
   • Display the annotated test image with bounding boxes and print instance data.

📚 Citation

This project is based on the SIFT feature descriptor introduced by David Lowe. Please cite the following paper if you use this work:

David G. Lowe. "Distinctive Image Features from Scale-Invariant Keypoints", International Journal of Computer Vision, 60, 91–110 (2004).
Springer Link

🪪 License

This project is licensed under the MIT License.