Foundations of Computer Vision

Course project of COMP4901L - a clone of CMU 16-385 course Team: Daniel Cheung, Dipsy Wong Repo: https://github.com/dipsywong98/COMP4901L

One of the hardest and most intensive UG Computer Science course in HKUST, 7 programming projects in 13 weeks. ~~Thanks god our super difficult final is cancelled~~.

Matlab warm-up
image filtering and hough transform
augmented reality and planar homography
3D reconstruction
physics-based vision
scene recognition with bag of words
digit recognition with CNN
tracking objects in videos

1. image filtering and hough transform

Pipeline that applies kernals and hough transform to detect straight line edges

original image
apply sobel operator
thresholding
hough transform
lines

2. augmented reality and planar homography

Feature detectors (FAST/ Harris corner detections), feature descriptors (BRIEF, filter bank), homography, RANSAC, Panorama

Extract and match the features ▲ Uploading file..._6nl6yh4f5
Compute the homography and map new image

Panorama left right result

3. 3D reconstruction

Sparse reconstruction ▲ Uploading file..._jip2c1p53

dense reconstruction

pose estimation

4. Physic based reconstruction

Given albedo and normal, render model under different light source

Given pictures under different light source, compute albedo and normal

5. scene recognition with bag of words

Given a set of filters, set of training images, sample some points using random or corners, compute their filter responses, k-mean cluster the responses as dictionary. Then map each pixel of all images into words in the dictionary, this word map then convert into a histogram counting the relative frequency of each word. Next use image feature distance/ NN/ kNN/ SVM to do the classification

original
filter response on LAB color system
collect sampling points
image to word map
classification kNN 57.5% accuracy, k=14 LibSVM 63.75% accuracy, HOG+SVM 80% accuracy