xiph
mv candidates in BMA for multi-ref prediction should only include the same ref mvs.
I found this bug in my bi-pred work, and expect fixing this bug will help improving the lost coding gain of muti-ref as well! Specifically, inside the function od_mv_est_init_mv(), whenever it construct mv candidates from spatial or temporal neighbors, it should only consider the same ref mvs.
Yes, I know the mv predictors is fixed before and I also use it. But I mean when deciding the candidates mvs in BMA, not the predictors, pred[] which returns from get_predictor() function. So, for ex, cneighbors[ci]->bma_mvs[0][ref], we don't know whether the position cneighbors[ci] will have valid mv for ref == backward prediction (or, golden frame), unless we check it thru mv_grid_pt[].ref.
Oh okay. Yeah I don't think reducing the # of candidates checked will actually be beneficial, but I'll leave it open if someone wants to try.
I hope this can fix my big loss of b-frame (bma only) for some sequences. FYI, when I applied your fix for get_predictor(), I got +5% for B-frame+bma only! So, I hope again excluding diffrent ref mvs will improve a lot.
Currently, I think the best fix would be to add ref field in od_mv_node, then look up it to check whether the candi mv has the same ref. Otherwise, use vx,vy position embedded in od_mv_node to access mv_grid_pt's ref field.
-yushin
On Mon, Oct 12, 2015 at 11:54 AM, Thomas Daede [email protected] wrote:
Oh okay. Yeah I don't think reducing the # of candidates checked will actually be beneficial, but I'll leave it open if someone wants to try.
— Reply to this email directly or view it on GitHub https://github.com/xiph/daala/issues/144#issuecomment-147490647.
Thanks! Yushin