Vinamra Singh's new entropy coding scheme beats Huffman in H.264/AVC CABAC
ECB achieves 0.113 bits/symbol savings on Kodak image benchmark over Huffman.
Vinamra Singh's new research introduces entropy-conserving binarization (ECB) into the CABAC entropy coder of H.264/AVC, the foundation for HEVC and VVC. ECB theoretically preserves entropy when mapping m-ary symbols to m-1 binary strings, but had never been tested inside a production binary arithmetic coder. Singh integrated ECB alongside three alternatives—UEG, single-context Huffman, and HuffmanPos (with per-bin-position contexts)—all sharing a single M-coder backend. Benchmarks covered synthetic sources, DCT residuals from a procedural image, and the full 24-image Kodak dataset (2,480 bit-exact, round-trip trials).
On the Kodak suite, ECB beat single-context Huffman at every Q, with the gap growing from 0.031 to 0.113 bits per symbol. However, HuffmanPos—using Huffman codewords but allocating one context per bin position—outperformed ECB on 12 of 15 source cells and lost by at most 0.56 percentage points on the other three. This shows that context allocation over the bin stream, not the binarization's per-symbol bin count, is the dominant driver at low source entropy. ECB's decoder latency penalty (7–10x on large alphabets) stems from its O(N*m) decoder; Singh sketches an interleaved single-pass variant to close this gap. All code, benchmarks, and raw data are open source.
- ECB beats single-context Huffman on the Kodak image benchmark at every Q value, saving 0.031–0.113 bits per symbol.
- Huffman with per-bin contexts (HuffmanPos) outperforms ECB on 12 of 15 source cells, highlighting context allocation over bin count as key.
- ECB decoder latency is 7–10x higher than alternatives due to O(N*m) algorithm; an interleaved single-pass fix is proposed.
Why It Matters
New entropy coding technique could improve video compression efficiency by up to 0.113 bits per symbol.