Simulation Results¶
We have integrated the learned ResNet model into HTM16.2 (which is
the reference software of 3D-HEVC).
Several simulations are carried out to further evaluate the performance of the proposed algorithm.
BD-BR and BD-PSNR metrics [REF2] are employed.
Simulation Environments¶
Device¶
- Macbook Pro (15-inch, Mid 2015)
- Processor 2.2GHz Intel Core i7
- Memory 16GB 1600MHz DDR3
- Nvidia GTX980, Memory 4GB (External GPU)
Video Sequences¶
Data are collected from four video sequences.
(This table is copied from Training Data Source)
# Name of the Sequence Resolution Usage Frames 1 Balloons 1024x768 train/test/validation 300 2 Kendo 1024x768 train/test/validation 300 3 PoznanStreet 1920x1088 train/test/validation 250 4 UndoDancer 1920x1088 train/test/validation 250
We want to make sure every sample that will be predicted has never been seen by the learned model. Otherwise it will be cheating.
Anther four sequences which have never been seen by the learned ResNet model are used for simulation:
# Name of the Sequence Resolution Usage Frames 1 Newspaper 1024x768 Simulation 300 2 GhostTownFly 1920x1088 Simulation 250 3 PoznanHall2 1920x1088 Simulation 200 4 Shark 1920x1088 Simulation 300
Configuration¶
The common test condition defined in [REF1] are used.
All the sequences are encoded as I-Frame.
Simulation Results¶
Figure 1. Time Saving for DMM1 Wedgelet Searching and Coding Performance of the Proposed Method
Figure 1. Time Saving for the total encoding process and Coding Performance of the Proposed Method