- Feature detection (implementation of Harris corner detection)
- Feature tracking : optical flow (implementation of Lucas-Kanada algorithm)
- Shape alignment (Affine transformation)
- Feature matching (Matchin SIFT descriptors using 2 metrics)
All the implementations are programmed from scratch without using any in-built functions except for minor operations such as convolution
We first compute the image gradients using which we form the second moment matrix. The harris score I have considered is the ratio of determinant of second moment matrix to trace of the matrix.
| Corner detection |
|---|
 |
After getting the key points in the image, we start to track them across a sequence of images. We do this by taking patches of images t and t+1 and solve for the displacement by least squares and add them to the tracking points.
| Second set of Tracked Points |
|---|
 |
| Progress for 20 points | Full sequence |
|---|---|
 |
 |
First, a transformation matrix is initialized and ICP (Iteratice Closest Point) alogrithm is implemented to compute affine transformation between two images. Image is transformed using this affine transformation matrix.
| Image 1 | Image 2 |
|---|---|
 |
 |
| Image 1 aligned as Image 2 |
|---|
 |
The SIFT descriptors are matched using two metrics. One is distance ratio of two nearest neighbors and the second one being distance of the nearest neighbhor
| Matches using only distance | Matches using distance ratio of two nearest neighbors |
|---|---|
 |
 |