Skip to content

Release V25.03

Pre-release
Pre-release

Choose a tag to compare

@djie1 djie1 released this 31 Mar 08:24
· 8 commits to master since this release

New and Changes in v25.03

  • Update iVSR plugion to FFmpeg n7.1. Note the options setting of "dnn_processing" filter is updated.
  • Enabled OpenVINO 2024.5 in the dockerfiles.
  • Added support of Rocky linux 9.3.
  • Enabled support of 10bit video in the iVSR FFmpeg plugin and SDK; SVP models updated to support 10bit video.
  • Enabled async processing mode in iVSR SDK and its plugin.
  • iVSR API performance improved by 3-4.5x (SVP-Basic model as inference workload), resulting in faster inference pipeline.
  • New SVP Y models introduced with performance improvement (1.58x for SVP-Basic, 8x for SVP-SE, on 1 Flex170 card).

Bug Fixes

  • Provided robust input resolution checking, and input options checking.
  • Fixed segment fault issue in vsr_sample if the reshape_values have odd number.
  • Fixed one random crash issue in infer_completion_callback() API.

Known Limitations/Issues

  • If the model-guard protected model is loaded, it prints the following error messages. They can be ignored as its normal procedure for this kind of model files.

    [libprotobuf ERROR thirdparty/protobuf/protobuf/src/google/protobuf/text_format.cc:335] Error parsing text-format tensorflow.GraphDef: 1:2: Message type "tensorflow.GraphDef" has no field named "T".

    [libprotobuf ERROR thirdparty/protobuf/protobuf/src/google/protobuf/text_format.cc:335] Error parsing text-format tensorflow.GraphDef: 1:2: Message type "tensorflow.GraphDef" has no field named "T".
  • For Enhanced EDSR, or SVP models, some resolutions like 240x426/144x256/234x426/238x246/220x400 don't work with OpenVINO 2023.2. You can switch to OpenVINO 2024.5.
  • GPU driver is not installed in the docker files for Rocky Linux 9.
  • There is color shift for some green/yellow colors after processing with Enhanced EDSR model.
  • Specifying "-vframes " in the ffmpeg command line to let pipeline exit before EOS, may cause core dump. You may remove this option, or specifying "nireq=1" option to dnn_processing filter to avoid it.