Vision-Project

1 Image Formation and Sensing

Color image demosaicing:

Implement the nearest neighbour, linear interpolation algorithm. Then, implement a more advanced method such as the adaptive gradient-based algorithm

2 Image Denoising

Image Denoising

Implement a mean filter Implement a mean filter using a filter matrix. Implement a Gaussian filter.

Edge Detection

Implement a delta filter, a Laplacian filter.

Hybrid Images

Fourier transform. Implement low and high pass filters and apply them to images. Create a hybrid image using high-pass and low-pass fitlered images.

3 Construct a panorama by stitching several individual and overlapping images

Homography

Warping

SIFT and RANSAC

Image Blending

Creating Panoramas

4

Implement NN Layers

Implement the backward_pass of fully connected layer Implement the backward_pass of convolutional layer

Fintuning and Adversarial Attacks

Implement the train function to complete fintuinig Adversarial attacks on 4 images of 4 classes

5 Stereo Depth

Greedy Stereo Matching

From the left and right image, calculate the disparity between each pixel using a greedy matching algorithm. You may assume that the stereo pairs are rectified, which means the camera pairs are only horizontally translated. In your PDF, be sure to include both the code and the visualization of the estimated disparity.

Distance Function: There are a variety of distance functions that you can use. For this problem, you can just use sum of squared differences between RGB patches.

dynamic programming based stereo estimation algorithm: Viterbi algorithm

implement a dynamic programming based stereo estimation algorithm that incorporates the spatial smoothness between adjacent disparities. For the local matching cost, you can use the same sum-of-squared-differences

6 Motion

Lucas-Kanade optical flow algorithm:

using the brightness constancy constraint to implement the Lucas-Kanade optical flow algorithm, and evaluate it on synthetic data