Banzai Beach v23.3 Pre-production Release

BanzaiBeach Pre-production Release Notes

Release Version 23.3

Release Information

This release includes components mainly related to Raisr only (Dataset preparation, training, testing, metrics calculation, comparison, visualization) and also supports default algorithms like bicubic, bilinear, lanczos for reference.
See README for more details.

Main Features

• Super Resolution algorithm dataset building, from both images and videos.
• RAISR (Rapid and Accurate Image Super Resolution) filter training, and testing.
• For comparison algorithms, it enables CV algorithms including bicubic, bilinear, and lanczos.
• It supports image quality metrics including SSIM, MS-SSIM, PSNR, VMAF, GMAF, and HAARPSI and supports logging and plot generation.
• Results comparison and visualization are supported to provide easy-to-use interfaces for users.
• It is a flexible framework that supports single-stage execution or multi-stage execution of the pipeline.
• This framework has specific support for both 8-bit & 10-bit images/videos.

Known Limitations

• Only mp4 videos are supported when using videos to build a dataset.
• Only image formats like png/jpg, etc. Can't support y4m/yuv files.
• Only Raisr training is supported in the training part.
• Downscale algorithm index have malposition, please set downscale_algorithm parameter in the correct index order when do downscaling:
  • 0=nearest, 1=bilinear, 2=bicubic, 3=area, 4=Lanczos, 5=Blur, 6=Random