Several 3D computer vision algorithms. Core modules are implemented without using OpenCV library. This project is from NTU 3D Computer Vision class assignment.
Create data folder and download data after clone this repository.
python3 3DCV/{task name}.py
Given three color images 1-0 (1-0.png), 1-1 (1-1.png), and 1-2 (1-2.png), please compute the homographies that warps the anchor image 1-0 to target image 1-1 and 1-2.
| image 1-0 | image 1-1 | image 1-2 |
|---|---|---|
 |
 |
 |
Result:
| # pairs | MSE (1-0, 1-1) | MSE (1-0, 1-2) |
|---|---|---|
| 4 | 21.6986 | 43.8561 |
| 8 | 0.2038 | 86.8156 |
| 20 | 0.0328 | 858.1783 |
| 80 | 0.0115 | 18.6487 |
Rectification is one of the most fundamental techniques when digitizing documents. Given an image of a document captured by the camera, please recover its original geometric property which is lost after perspective transformation. The following figure is an example of rectifying a photo of a book.
Result:
For each validation image, compute its camera pose with respect to world coordinate. Find the 2D-3D correspondence by descriptor matching, and solve the camera pose.
Result:
For each camera pose you calculated, plot the trajectory and camera poses along with
3d point cloud model. Provide some discussion on the visualized results.
Implement a VO based on two-view epipolar geometry.
