Name: Tzu-Chuan Lin
Rectified:
Rectified:
These are the pictures I have taken:
Combined:
Q: Whats the most important/coolest thing you have learned from this part?
cv2.findHomography
.In my implementation, I just directly tranformed an image into gray scale and then performed the edge detection.
The result of harris corner detection + ANMS(Adaptive Non-Maximal Suppression)
NOTE: Red points are the points still there after ANMS.
The feature descriptors (before normalization):
NOTE: Because my images are with high-resolution, I set the patch size be 64x64
(resized from 128x128
patch) to increase the descriptiveness of each patch.
I used SSD (i.e.np.sum((img1-img2)**2)
) to measure the similarity between feature two descriptors.
In my implementation, I gave RANSAC 1000 iterations.
PartA (manual labeling) | PartB (automatic pairing) |
---|---|
PartA (manual labeling) | PartB (automatic pairing) |
---|---|
PartA (manual labeling) | PartB (automatic pairing) |
---|---|
Detail: I derive the cylindrical mapping by hand and use the pyramid search + SSD to find the best alignment for each (i-1, i)
pair in my images.
See: here for the original images.
Sproul Plaza:
NOTE: Markdown does not allow me to display too long image, so you might want to directly click into that to see the full image.
However, you can notice some ghosting inside the image. I think it may be caused by inaccruate focal length or because I used a tripod that cannot be rotated horizontally.
I rotated one image like this:
But I can still get the same result as the image above.
(I do not provide the result here again because it is almost the same)