We discussed the quality of superpixels here:
We brainstormed a few heuristics to improve superpixel segmentation and decided on the following ideas:
1- Get rid of low quality images
2- Remove compression noise using median filter
3- Use other superpixel algorithms such as TurboPixels
4- See if Jue's paper provides auxiliary data for text pixels
5- Heuristics to use text annotations to detect pixels belonging to the text
6- Hand annotate which superpixels must go together.
We noted that how superpixels relate to eachother matters. Probably we need various kinds of features including color, shape and stroke width transform feature. We also discussed ideas to use VOC style detectors and RCNN features but we planned to study them later on.
* We also talked about using the manual bounding boxes to group the superpixels, to see if the result looks reasonable and can be used as training data.
ReplyDelete* We discussed a background model, by k-means (k=2) heavily weighted to sampling the image edges.
* We discussed adding stroke width transform (SWT) to the rgbxy mean shift feature vector, as well as distances to background models formed above.
We also discussed some some heuristics that can be used to group superpixels together:
a. Being separated by background-like pixels.
b. Similarity in color and texture
Rather than k-means, we might want to find the background color by using the most frequent color in the image (perhaps weighted toward edges). Perhaps weighted by whether you're in a superpixel that touches the image boundary.
ReplyDelete