JVET-D0102 EE3: Generalized bi-prediction [C.-C. Chen, X. Xiu, Y. He, Y. Ye (InterDigital)]

In the 3rd JVET meeting, the Exploration Experiment 3 was established in JVET-C1011 to study the generalized bi-prediction (GBi) technique and report its coding performance on top of the JEM-3.0 software platform. GBi is an inter prediction mode that allows bi-prediction weights to be signalled at the leaf node of the quadtree plus binary tree structure and applied to each sample in bi-prediction. This contribution is based on JVET-C0047 (i.e. the original proposal of GBi) to evaluate the coding performance of GBi under three different settings for its weights: Test 1 with {3/8,1/2, 5/8}, Test 2 with {1/4, 3/8, 1/2, 5/8, 3/4} and Test 3 with {−1/4, 1/4, 3/8, 1/2, 5/8, 3/4, 5/4}. The coding results of each test shown in the following are reported relative to the JEM-3.0 anchor, where the metrics in use are based on average Y BD-rate reduction, encoding time and decoding time.

(Test 1) Random Access: 0.6%, Low Delay B: 0.4%, Encoding Time: 126%, Decoding Time: 99%;

(Test 2) Random Access: 0.7%, Low Delay B: 0.5%, Encoding Time: 154%, Decoding Time: 99%;

(Test 3) Random Access: 0.8%, Low Delay B: 0.9%, Encoding Time: 186%, Decoding Time: 99%.