Weekly meeting: 22 Jun 2012
1. Status updates
Tim: General clean-ups and entropy coder
+ Allow encoder buffer resizing [done]
+ Convert to multiply-free approach [done, except for the bugs]
+ Reduce arithmetic to 16 bits [done, except for the bugs]
Jean-Marc:
+ PVQ encoder (2 new schemes tested)
- Original scheme: one symbol up to last non-zero value (enc+dec)
- New scheme 1: one symbol per pulse (distance to next pulse w/Laplace distribution) (enc+dec)
- New scheme 2: splits into chunks and uses CWRS (simulation only, 1% penalty @ N=128)
- Almost all roughly equal, no big gains without something radical
+ Intra mode coder experiments
+ Comments/Doxygen on the existing code
+ Generic coef encoder
Nathan:
+ Completed exhaustive search 8x16 R=t,D=t where CG>=9.2 (68139 transforms)
+ Updated Matlab code to process all block modes
+ Started on k-means in Java using subset1-y4m
- look at average residual error (L2 error as we drop multiplies)
+ Still need to look at Type III
Greg:
+ Integration of general value coder so all data is coded
2. Next Steps
Tim:
+ Finish new DCT factorizations
- This should unblock looking at larger transform sizes and block size switching
+ Complete SPIHT-like coder (so we can compare with PVQ coefficient coders)
Jean-Marc:
+ Continue INTRA mode coding (RDO+probability modeling)
+ Doing some RDO for the PVQ stuff
Greg:
+ Commit, expose r/d biases for JM
+ visualizations of the 8x16 prefilter metrics
http://people.xiph.org/~unlord/8x16rd_all.tar.gz
+ 4x8 prefilter results on real images (need to figure out scaling)
Nathan:
+ Run 16x32 R=t,D=t on kindle01 using solver
-- dump of https://daala.etherpad.mozilla.org/2 on 20120629.