clashroyaleisgood / course_computervision

電腦視覺(英文授課) Computer Vision, in NCTU

Course_ComputerVision

電腦視覺(英文授課) Computer Vision, in NCTU

There are 3 homeworks(projects)

• HW1: Photometric Stereo
• HW2: Image Stitching
• Final Project: Depth Estimation from Stereo Images

Environment

dependency:
python==3.6.13

matplotlib==3.3.4
numpy==1.19.5
open3d==0.15.1
opencv-python==4.5.5.62
scipy==1.5.4
tqdm==4.63.0

maybe that's all(?

HW1: Photometric Stereo

Introduction:

Given many image taken(rendered) from same viewpoint with different light directions(LightSource.txt).
Try to estimate the depth(height).

pic1	pic2	pic3	...
			...

Result:

Detail:

Use Diffuse reflection method to simulate the image color.
Calculate it backward to get the height.

$I_{diffuse} = consant * (n \cdot l)$
$n: \text{normal vector}$
$l: \text{light direction vector}$

1. Estimate every normal vector in image[i, j].
   $l = norm(I_{diffuse} \times l_{inv})$
   $l_{inv}: inverse(l)$
2. Estimate every Gradient vector in (x, y) directions in image[i, j]
   by normal vectors.
3. Estimate the surface height by each Gradient vector.

HW2: Image Stitching

Introduction:

Stitching many images

pic1	pic2

Result:

Detail:

1. Use SIFT to find features in two images.
2. Find the most similar feature pairs $P(p_i, p_j)\ \forall\ p_i \in I_1$
   by kNN algorithm(k == 1), and Lowe's Ratio test
   where:
   $I_1: \text{image1}$
   $p_i: \text{the location of the features that SIFT find in }I_1$
   $\text{same as } p_j \text{ to } I_2$
3. Find the best Homography Matrix that transform image 1 to image 2 viewpoints by RANSAC algorithm.

   RANSAC
   for try_many_times:
       locatoins_of_4_pairs = randomly_get_4_pair()
       matrix = calculate_homography(locatoins_of_4_pairs)
   
       fit_result = try_fit(matrix, location_of_every_image1_feature)
       score = count_matched(fit_result, location_of_paired_image2_feature, threshold)
   
       if score > max_score:
           best_matrix = matrix
   

Final Project: Depth Estimation from Stereo Images

Introduction:

Estimate the depth from two images which is taken by same camera, with same direction but different horizontal location.
But I also use image Rectification to rectify two images that have differenct camera locatoin or direction into same direction and horizental movement only.

Result:

just see: https://github.com/clashroyaleisgood/Course_ComputerVision/tree/main/Final_Project

Detail:

just see: https://github.com/clashroyaleisgood/Course_ComputerVision/tree/main/Final_Project