My median filter gets the disparity map estimated by Census Transform costs ( hamming distance for 8-bit strings) worse by around %5 of the original error. Unlike, Rank transform based disparity matching; which shows enhancement in the error results between ground truth and the output disparity map after applying median filtering. What could be the reason behind this?
- No sub-pixel estimation is done
- Census transform has a size of 3 x 3 window
- Search size is pixel by pixel ( each pixel is 8-bit string in this case)
- Disparity range is between 0 and 255
the best pixel location will be at the location has the minimum hamming distance
hamming distance = sum(First_string xor Second_string )
- so min h_d is 0 and max h_d is 8
https://books.google.iq/books?id=CKxnCgAAQBAJ&pg=PA310&lpg=PA310&dq=Why+is+not+median+filter+compatible+with+Census+transform+based+disparity+matching?&source=bl&ots=MUs4hZd3hq&sig=ACfU3U2rFUhkHWF0kQPQRYfrkvSOEqMQ-w&hl=ar&sa=X&ved=2ahUKEwixs_eRp8LmAhXfSxUIHTbdBBUQ6AEwAHoECAkQAQ
Try singular values decomposition average mean or PCA instead median filter.
Muhammad Ali
Thank you for your suggestions, if they help in MATLAB it would be great. However, I think they should be much more costly at hardware implementations.
https://books.google.iq/books?id=CKxnCgAAQBAJ&pg=PA310&lpg=PA310&dq=Why+is+not+median+filter+compatible+with+Census+transform+based+disparity+matching?&source=bl&ots=MUs4hZd3hq&sig=ACfU3U2rFUhkHWF0kQPQRYfrkvSOEqMQ-w&hl=ar&sa=X&ved=2ahUKEwixs_eRp8LmAhXfSxUIHTbdBBUQ6AEwAHoECAkQAQ
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing