New features
- Updated VMAF algorithm to v0.6.1, including:
- Added a custom model for cellular phone screen viewing
- Trained using new dataset, covering more difficult content
- Elementary metric fixes: ADM behavior at near-black frames, motion behavior at scene boundaries
- Compressed quality score range by 20% to accommodate higher dynamic range
- Use MLE instead of DMOS as subjective model
- Added command line ffmpeg2vmaf, which takes encoded videos (instead of raw YUV) as input.
- Allow specifying crop and pad parameter in dataset files.
- Speeded up VMAF convolution operation by AVX.
- Added Travis continuous integration.
- Add implementation of KFLK - quality metric evaluation method based on AUC. Refer to: L. Krasula, K. Fliegel, P. Le Callet, M.Klima, "On the accuracy of objective image and video quality models: New methodology for performance evaluation", QoMEX 2016.
- Add options to use custom subjective models in run_vmaf_training and run_testing commands.
Bug fixes
- Revamped process-level parallelization for Executor.
- Fixed vmafossexec memory leakage.
- Fixed command line run_testing issue. Add command line test cases.
- Fixed a bug in DatasetReader.to_aggregated_dataset_file.
- Issue #36: SSIM and MS-SSIM sometimes get negative values.