Clustering-Deepening: Clustering Algorithms for Object Tracking & Image Segmentation
Clustering-Deepening is an educational and experimental project that implements, analyses, and compares a wide spectrum of clustering algorithms. Beyond benchmarks, each algorithm is applied to two concrete vision tasks: object tracking and image segmentation.
Algorithms Covered
| Algorithm | Key Properties |
|---|---|
| k-Means | Centroid-based, Euclidean, fast |
| k-Means++ | Improved initialisation, more stable |
| k-Medians | Median centroid, robust to outliers |
| k-Medoids | Cluster centres must be data points |
| Mean-Shift | Non-parametric, bandwidth-driven |
| DBSCAN | Density-based, arbitrary shapes, detects noise |
| GMM (EM) | Probabilistic, soft assignments |
Evaluation Metrics
- Davies-Bouldin Index: lower = more compact, well-separated clusters.
- Rand Index: agreement with ground-truth labels (where available).
- Silhouette score for internal evaluation without labels.
Applications
- Object Tracking: use clustering to group detections across frames, associating each cluster with a tracked object.
- Image Segmentation: apply colour-space clustering to partition images into semantically meaningful regions without supervision.
Technology
Python, scikit-learn, NumPy, OpenCV, Matplotlib, Jupyter Notebooks.