This artificial test image (from the Dava engine) is tough for PVRTC. Left is original, middle is my encoding (44.3 Y dB), right is PVRTexTool (37.89). Those artifacts on the 1 and 0 are so annoying.pic.twitter.com/S3dyfLod4e
আপনি আপনার টুইটগুলিতে ওয়েব থেকে এবং তৃতীয়-পক্ষ অ্যাপ্লিকেশনগুলির মাধ্যমে অবস্থান তথ্য যেমন শহর বা সুনির্দিষ্ট অবস্থান যোগ করতে পারবেন। আপনার কাছে আপনার টুইটের অবস্থান ইতিহাস মোছার বিকল্প থাকবে। আরও জানুন
IMO, it's an essential step with "graphics" or alpha channel imagery. You can't rely on the local, main axes being aligned with, say, luma. FWIW, PVRTexTool does a low-pass filter to separate low and high frequency data, and does a weighted PCA on the high-frequency...
...data regions. Like BT1 encoding, the colours can be projected on to the axis & tuned to come up with end points. The trouble is eigenvector analysis doesn't care if it gives you {r,g,b,a} or {-r,-g,-b,-a}, as the principal vector, so they can potentially flip from region to...
I remember this algorithm in one of the early PVRTC papers. I have an implementation somewhere. I'll give it a try again.
Right now I compute each 4x4 block's RGBA AABB and floor/ceil it to compute the initial block L/H endpoints (like Lim's PvrTcCompressor). They are then refined using 1x1 block least squares and table lookups. These tricky "mod flips" are the last remaining major artifact to fix.
Interesting. I suspect that would be easy enough to modify in PVRTextool as the filter weights for the initial guesses are table driven... just need to add in a lot of zeros.
Lin's approach is very "stable" and is useful in real-time encoders. High frequency image features get reliably smoothed out. (Its output is basically reliably crappy!) It's usually a good source for further refinement (using LS or tables). But it suffers from mod-flips.
A few years back, a team member was working on a GPU compressor for PVRTC but it got shelved (well, pushed off the back burner and into the freezer). It really needed a way of "type punning" the texture format and (at the time) this wan't available.
Mind you, I don't think it had a way of dealing with the alignment problem either: Easiest to do that on the CPU.
Seems like the map coloring problem with three colors...
Andrew sent me his improved PVRTC encoding. So good I did a doubletake.pic.twitter.com/qvZzwgU19j
টুইটার তার ক্ষমতার বাইরে চলে গেছে বা কোনো সাময়িক সমস্যার সম্মুখীন হয়েছে আবার চেষ্টা করুন বা আরও তথ্যের জন্য টুইটারের স্থিতি দেখুন।