MY WORKS - Spring '17 |
|
Image I/O & LibrariesColor image has more total number of pixels than a grayscale image. This is because the color image has three channels of color (red, green, blue) while a grayscale image only has one layer.
To convert the grayscale image, check every single pixel of the image to find all the pixels that is greater than 128. Then, change these pixels to 255 (white color).
The overflow issue happens when the input image arrays and output array are not typecasted properly which results in the resulting data for the output array is bigger than the upper bound of the datatype that would store it (uint8, in this case).
PanoramasChanging the number of matched features causes my input images to be swapped and put together differently. If there are too few matches, the images swapped strangely. The overlap section of the pairs of images are not calculated the way the output should look like, thus, the images are not stitched right.
Big images (e.g: 5324 x 4200) do not warp or blend well. The images need to have almost similar or light change of lighting near the seam area for the output to look natural.
|
Epsilon PhotographyOne chess piece moves at a small space with time while other chess pieces stay fixed.
BlendingPurpose of blend() is to form a combined pyramid by using a Gaussian pyramid mask a weight.
Photo of spaceOpenSFM takes very long time to generate point cloud with about 200 photos (each one 640x480). It also difficult to explore around the point cloud after it is generated. To look at a specific image cluster, I had to move my mouse very slowly and in small movement to get to the area I wanted.
Video TexturesI determined the best alpha value for each video texture by looking at their transition matrix. The dark areas provide information of the location of the best transitions between end and start frames. If the output loop is longer in length and missed the a shortest loop (normally have better transition), then I know I should decrease the alpha value. In contrast, I will want to increase the alpha value to get the longer potential loop.
As alpha increases, it gives more weight to the loop’s length. On the other hand, decreasing alpha gives more weight to the transition. Any alpha change will affect the scores.
|
Gradients & EdgesBox Filter:
Features detection & Matching
I chose scenes that did not have similar color or figures with my test object. This way, features matching would not be confused by the surrounding and mismatch my test object’s features with other similar pattern/shape in the same background.
I also tried plain background, different capture distance from the test object, and different light time of the day. §Plain background minimized the maximum distraction around my test object. Distance adjustment and different light time give different match results. |