- Divides image into localized regions
- Generates histogram for each localized region
- Histograms are created using gradients and orientation of pixel intensities
- In this way, different shape accounts for different histogram
- How it works:
1. Convert image to grayscale
2. calculate magnitude and angle of gradients (you may use sobel)
3. Divide images into grid-cells
4. For each cell, calculate the histogram of gradient direction
5. Block-normalize cells to reduce imabalanced highlights / shadows
6. Use this combination as feature vector