1 =============================================
2 SNOW Video Codec Specification Draft 20070103
3 =============================================
7 This Specification describes the snow syntax and semmantics as well as
9 The decoding process is precissely described and any compliant decoder
10 MUST produce the exactly same output for a spec conformant snow stream.
11 For encoding though any process which generates a stream compliant to
12 the syntactical and semmantical requirements and which is decodeable by
13 the process described in this spec shall be considered a conformant
19 MUST the specific part must be done to conform to this standard
20 SHOULD it is recommended to be done that way, but not strictly required
22 ilog2(x) is the rounded down logarithm of x with basis 2
29 u unsigned scalar value range coded
30 s signed scalar value range coded
43 if(keyframe || always_reset)
46 version u header_state
47 always_reset b header_state
48 temporal_decomposition_type u header_state
49 temporal_decomposition_count u header_state
50 spatial_decomposition_count u header_state
51 colorspace_type u header_state
52 chroma_h_shift u header_state
53 chroma_v_shift u header_state
54 spatial_scalability b header_state
55 max_ref_frames-1 u header_state
59 update_mc b header_state
61 for(plane=0; plane<2; plane++){
62 diag_mc b header_state
63 htaps/2-1 u header_state
64 for(i= p->htaps/2; i; i--)
65 |hcoeff[i]| u header_state
68 update_qlogs b header_state
70 spatial_decomposition_count u header_state
75 spatial_decomposition_type s header_state
77 mv_scale s header_state
79 block_max_depth s header_state
82 for(plane=0; plane<2; plane++){
83 quant_table[plane][0][0] s header_state
84 for(level=0; level < spatial_decomposition_count; level++){
85 quant_table[plane][level][1]s header_state
86 quant_table[plane][level][3]s header_state
94 for(y=0; y<block_count_vertical; y++)
95 for(x=0; x<block_count_horizontal; x++)
99 mvx_diff=mvy_diff=y_diff=cb_diff=cr_diff=0
103 if(level!=max_block_depth){
104 s_context= 2*left->level + 2*top->level + topleft->level + topright->level
105 leaf b block_state[4 + s_context]
107 if(level==max_block_depth || leaf){
108 intra b block_state[1 + left->intra + top->intra]
110 y_diff s block_state[32]
111 cb_diff s block_state[64]
112 cr_diff s block_state[96]
114 ref_context= ilog2(2*left->ref) + ilog2(2*top->ref)
116 ref u block_state[128 + 1024 + 32*ref_context]
117 mx_context= ilog2(2*abs(left->mx - top->mx))
118 my_context= ilog2(2*abs(left->my - top->my))
119 mvx_diff s block_state[128 + 32*(mx_context + 16*!!ref)]
120 mvy_diff s block_state[128 + 32*(my_context + 16*!!ref)]
137 for(level=0; level<spatial_decomposition_count; level++){
155 this MUST NOT change within a bitstream
158 if 1 then the range coder contexts will be reset after each frame
160 temporal_decomposition_type
163 temporal_decomposition_count
166 spatial_decomposition_count
171 this MUST NOT change within a bitstream
174 log2(luma.width / chroma.width)
175 this MUST NOT change within a bitstream
178 log2(luma.height / chroma.height)
179 this MUST NOT change within a bitstream
185 maximum number of reference frames
186 this MUST NOT change within a bitstream
189 indicates that motion compensation filter parameters are stored in the
193 flag to enable faster diagonal interpolation
194 this SHOULD be 1 unless it turns out to be covered by a valid patent
197 number of half pel interpolation filter taps, MUST be even, >0 and <10
200 half pel interpolation filter coefficients, hcoeff[0] are the 2 middle
201 coefficients [1] are the next outer ones and so on, resulting in a filter
202 like: ...eff[2], hcoeff[1], hcoeff[0], hcoeff[0], hcoeff[1], hcoeff[2] ...
203 the sign of the coefficients is not explicitly stored but alternates
204 after each coeff and coeff[0] is positive, so ...,+,-,+,-,+,+,-,+,-,+,...
205 hcoeff[0] is not explicitly stored but found by subtracting the sum
206 of all stored coefficients with signs from 32
207 hcoeff[0]= 32 - hcoeff[1] - hcoeff[2] - ...
208 a good choice for hcoeff and htaps is
211 an alternative which requires more computations at both encoder and
212 decoder side and may or may not be better is
218 minimum of the number of available reference frames and max_ref_frames
219 for example the first frame after a key frame always has ref_frames=1
221 spatial_decomposition_type
223 0 is a 9/7 symmetric compact integer wavelet
224 1 is a 5/3 symmetric compact integer wavelet
226 stored as delta from last, last is reset to 0 if always_reset || keyframe
229 quality (logarthmic quantizer scale)
230 stored as delta from last, last is reset to 0 if always_reset || keyframe
233 stored as delta from last, last is reset to 0 if always_reset || keyframe
234 FIXME check that everything works fine if this changes between frames
238 stored as delta from last, last is reset to 0 if always_reset || keyframe
241 maximum depth of the block tree
242 stored as delta from last, last is reset to 0 if always_reset || keyframe
248 Highlevel bitstream structure:
249 =============================
250 --------------------------------------------
252 --------------------------------------------
253 | ------------------------------------ |
257 | | ......... intra? | |
258 | | : Block01 : yes no | |
259 | | : Block02 : ....... .......... | |
260 | | : Block03 : : y DC : : ref index: | |
261 | | : Block04 : : cb DC : : motion x : | |
262 | | ......... : cr DC : : motion y : | |
263 | | ....... .......... | |
264 | ------------------------------------ |
265 | ------------------------------------ |
268 --------------------------------------------
269 | ------------ ------------ ------------ |
270 || Y subbands | | Cb subbands| | Cr subbands||
271 || --- --- | | --- --- | | --- --- ||
272 || |LL0||HL0| | | |LL0||HL0| | | |LL0||HL0| ||
273 || --- --- | | --- --- | | --- --- ||
274 || --- --- | | --- --- | | --- --- ||
275 || |LH0||HH0| | | |LH0||HH0| | | |LH0||HH0| ||
276 || --- --- | | --- --- | | --- --- ||
277 || --- --- | | --- --- | | --- --- ||
278 || |HL1||LH1| | | |HL1||LH1| | | |HL1||LH1| ||
279 || --- --- | | --- --- | | --- --- ||
280 || --- --- | | --- --- | | --- --- ||
281 || |HH1||HL2| | | |HH1||HL2| | | |HH1||HL2| ||
282 || ... | | ... | | ... ||
283 | ------------ ------------ ------------ |
284 --------------------------------------------
295 | | LL0 subband prediction
298 ------------------- \ |
299 | Reference frames | \ IDWT
300 | ------- ------- | Motion \ |
301 ||Frame 0| |Frame 1|| Compensation . OBMC v -------
302 | ------- ------- | --------------. \------> + --->|Frame n|-->output
303 | ------- ------- | -------
304 ||Frame 2| |Frame 3||<----------------------------------/
315 left and top are set to the respective blocks unless they are outside of
316 the image in which case they are set to the Null block
318 top-left is set to the top left block unless it is outside of the image in
319 which case it is set to the left block
321 if this block has no larger parent block or it is at the left side of its
322 parent block and the top right block is not outside of the image then the
323 top right block is used for top-right else the top-left block is used
327 level, ref, mx and my are 0
330 Motion Vector Prediction:
331 =========================
332 1. the motion vectors of all the neighboring blocks are scaled to
333 compensate for the difference of reference frames
335 scaled_mv= (mv * (256 * (current_reference+1) / (mv.reference+1)) + 128)>>8
337 2. the median of the scaled left, top and top-right vectors is used as
338 motion vector prediction
340 3. the used motion vector is the sum of the predictor and
341 (mvx_diff, mvy_diff)*mv_scale
345 ======================
346 the luma and chroma values of the left block are used as predictors
348 the used luma and chroma is the sum of the predictor and y_diff, cb_diff, cr_diff
349 to reverse this in the decoder apply the following:
350 block[y][x].dc[0] = block[y][x-1].dc[0] + y_diff;
351 block[y][x].dc[1] = block[y][x-1].dc[1] + cb_diff;
352 block[y][x].dc[2] = block[y][x-1].dc[2] + cr_diff;
353 block[*][-1].dc[*]= 128;
359 Halfpel interpolation:
360 ----------------------
361 halfpel interpolation is done by convolution with the halfpel filter stored
364 horizontal halfpel samples are found by
365 H1[y][x] = hcoeff[0]*(F[y][x ] + F[y][x+1])
366 + hcoeff[1]*(F[y][x-1] + F[y][x+2])
367 + hcoeff[2]*(F[y][x-2] + F[y][x+3])
369 h1[y][x] = (H1[y][x] + 32)>>6;
371 vertical halfpel samples are found by
372 H2[y][x] = hcoeff[0]*(F[y ][x] + F[y+1][x])
373 + hcoeff[1]*(F[y-1][x] + F[y+2][x])
375 h2[y][x] = (H2[y][x] + 32)>>6;
377 vertical+horizontal halfpel samples are found by
378 H3[y][x] = hcoeff[0]*(H2[y][x ] + H2[y][x+1])
379 + hcoeff[1]*(H2[y][x-1] + H2[y][x+2])
381 H3[y][x] = hcoeff[0]*(H1[y ][x] + H1[y+1][x])
382 + hcoeff[1]*(H1[y+1][x] + H1[y+2][x])
384 h3[y][x] = (H3[y][x] + 2048)>>12;
395 F-------F-------F-> H1<-F-------F-------F
399 F-------F-------F-> H1<-F-------F-------F
410 unavailable fullpel samples (outside the picture for example) shall be equal
411 to the closest available fullpel sample
414 Smaller pel interpolation:
415 --------------------------
416 if diag_mc is set then points which lie on a line between 2 vertically,
417 horiziontally or diagonally adjacent halfpel points shall be interpolated
418 linearls with rounding to nearest and halfway values rounded up.
419 points which lie on 2 diagonals at the same time should only use the one
420 diagonal not containing the fullpel point
424 F-->O---q---O<--h1->O---q---O<--F
432 h2-->O---q---O<--h3->O---q---O<--h2
440 F-->O---q---O<--h1->O---q---O<--F
444 the remaining points shall be bilinearly interpolated from the
445 up to 4 surrounding halfpel and fullpel points, again rounding should be to
446 nearest and halfway values rounded up
448 compliant snow decoders MUST support 1-1/8 pel luma and 1/2-1/16 pel chroma
449 interpolation at least
452 Overlapped block motion compensation:
453 -------------------------------------
458 Each sample in the LL0 subband is predicted by the median of the left, top and
459 left+top-topleft samples, samples outside the subband shall be considered to
460 be 0. To reverse this prediction in the decoder apply the following.
461 for(y=0; y<height; y++){
462 for(x=0; x<width; x++){
463 sample[y][x] += median(sample[y-1][x],
465 sample[y-1][x]+sample[y][x-1]-sample[y-1][x-1]);
468 sample[-1][*]=sample[*][-1]= 0;
469 width,height here are the width and height of the LL0 subband not of the final
480 Snow supports 2 wavelet transforms, the symmetric biorthogonal 5/3 integer
481 transform and a integer approximation of the symmetric biorthogonal 9/7
484 2D IDWT (inverse discrete wavelet transform)
485 --------------------------------------------
486 The 2D IDWT applies a 2D filter recursively, each time combining the
487 4 lowest frequency subbands into a single subband until only 1 subband
489 The 2D filter is done by first applying a 1D filter in the vertical direction
490 and then applying it in the horizontal one.
491 --------------- --------------- --------------- ---------------
492 |LL0|HL0| | | | | | | | | | | |
493 |---+---| HL1 | | L0|H0 | HL1 | | LL1 | HL1 | | | |
494 |LH0|HH0| | | | | | | | | | | |
495 |-------+-------|->|-------+-------|->|-------+-------|->| L1 | H1 |->...
496 | | | | | | | | | | | |
497 | LH1 | HH1 | | LH1 | HH1 | | LH1 | HH1 | | | |
498 | | | | | | | | | | | |
499 --------------- --------------- --------------- ---------------
504 1. interleave the samples of the low and high frequency subbands like
505 s={L0, H0, L1, H1, L2, H2, L3, H3, ... }
506 note, this can end with a L or a H, the number of elements shall be w
507 s[-1] shall be considered equivalent to s[1 ]
508 s[w ] shall be considered equivalent to s[w-2]
510 2. perform the lifting steps in order as described below
513 1. s[i] -= (s[i-1] + s[i+1] + 2)>>2; for all even i < w
514 2. s[i] += (s[i-1] + s[i+1] )>>1; for all odd i < w
516 \ | /|\ | /|\ | /|\ | /|\
517 \|/ | \|/ | \|/ | \|/ |
519 /|\ | /|\ | /|\ | /|\ |
520 / | \|/ | \|/ | \|/ | \|/
524 snows 9/7 Integer filter:
525 1. s[i] -= (3*(s[i-1] + s[i+1]) + 4)>>3; for all even i < w
526 2. s[i] -= s[i-1] + s[i+1] ; for all odd i < w
527 3. s[i] += ( s[i-1] + s[i+1] + 4*s[i] + 8)>>4; for all even i < w
528 4. s[i] += (3*(s[i-1] + s[i+1]) )>>1; for all odd i < w
530 \ | /|\ | /|\ | /|\ | /|\
531 \|/ | \|/ | \|/ | \|/ |
533 /|\ | /|\ | /|\ | /|\ |
534 / | \|/ | \|/ | \|/ | \|/
535 (| + (| + (| + (| + -1
536 \ + /|\ + /|\ + /|\ + /|\ +1/4
537 \|/ | \|/ | \|/ | \|/ |
538 + | + | + | + | +1/16
539 /|\ | /|\ | /|\ | /|\ |
540 / | \|/ | \|/ | \|/ | \|/
544 following are exactly identical
545 (3a)>>1 == a + (a>>1)
546 (a + 4b + 8)>>4 == ((a>>2) + b + 2)>>2
548 16bit implementation note:
549 The IDWT can be implemented with 16bits, but this requires some care to
550 prevent overflows, the following list, lists the minimum number of bits needed
553 A= s[i-1] + s[i+1] 16bit
558 s[i-1] + s[i+1] 17bit
561 3*(s[i-1] + s[i+1]) 17bit
567 finetune initial contexts
569 try to use the wavelet transformed predicted image (motion compensated image) as context for coding the residual coefficients
570 try the MV length as context for coding the residual coefficients
571 use extradata for stuff which is in the keyframes now?
572 the MV median predictor is patented IIRC
573 implement per picture halfpel interpolation
574 try different range coder state transition tables for different contexts
577 compare the 6 tap and 8 tap hpel filters (psnr/bitrate and subjective quality)
578 spatial_scalability b vs u (!= 0 breaks syntax anyway so we can add a u later)
589 GPL + GFDL + whatever is needed to make this a RFC