LSH-based Motion Estimation Alex Giladi Motion Estimation Motivation Utilize temporal redundancy for better video compression Improve video quality in MPEG-1 / MPEG-2 / MPEG-4 / H.264 AVC Definition For each block bs I(t), find closest br I(t-1) Objective: minimize the residue, bs – br Search ranges: ± 64 is common for NTSC broadcast (720x480) Assume: 16x16 blocks, done in MSE sense Algorithms Full search (brute force): Fast algorithms exist 7/26/2016 400K MSE computations for ± 16 on CIF (352x288) 132M MSE computations for ± 64 on 1080i (1920x1080) Real time: only 33.36ms per picture Speedups to full search Variants of logarithmic search Hierarchical motion estimation LSH-based motion estimation 2 Motion estimation and LSH Re-definition: 16x16 block is represented as a vector in 256. Similarity measure: L2 norm Hash functions: dot product with a random vector in 256 Algorithm Hash similar blocks to same buckets. Pick blocks that hashed to the same bucket Find best match among these 7/26/2016 LSH-based motion estimation 3 Motion estimation and LSH Why LSH? We don’t need exact answer We can allow longer pre-processing time We don’t need an answer where blocks are dissimilar An approximation is enough; These would not be coded using temporal prediction Very large search ranges can be supported Problems Search radius 7/26/2016 Large radius – too many candidate blocks to be considered Small radius – too many blocks have no pairs Requires a large amount of additional memory (vs. none in the regular algorithms) Requires several dot product computations per pixel LSH-based motion estimation 4 Results Answers are sufficiently close to the true values Complexity is reduced No answer for several blocks 7/26/2016 L2 distance Blocks found Additional MSE computations 128 87% 1075 64 80% 562 LSH-based motion estimation 5 Conclusion Using LSH Needs tuning and speedups Can potentially reduce ME complexity, when large search range is required. Extensions: Prefer closer vectors closer pictures represent x,y,t vector components in 259 Multiple references Fast candidate elimination techniques 7/26/2016 LSH-based motion estimation 6