Release notes version 08.04.00¶
New in this Release¶
SDK comes pre-installed with required dependencies and demo applications
Enable zero-copy in between GStreamer and runtimes
Support DMABUF import for v4l2 raw camera sources using tiovxisp element
Cross compile support for all EdgeAI repos
Support for separately enabling or disabling DL runtime
Support for streaming the demo output over the network
Added scripts to report stats on core loading and GStreamer pipeline
Support for SDE and DOF GStreamer plugins
Fixed in this Release¶
ADASVISION-5635 - tiovxisp output is wrong with dmabuf-import when sensor resolution is not multiple of 64
ADASVISION-5584 - Edge AI demo: Image classification demo freezes around 98~99 frames
ADASVISION-5583 - Edge AI demo: file input in multi-output demo causes demo to freeze
ADASVISION-5494 - Add support for simultaneous use of RPi camera and Rpi extension header
ADASVISION-5475 - J7 SK: The setup_script.sh fails when the board is configured in AP mode, which is the default with WiFi only case
ADASVISION-5463 - ONR-OD-8000 model fails with IMX390 in CPP app due to 1200x1200 resolution
ADASVISION-5349 - Setup script fails building without internet connection
ADASVISION-4997 - Multi-input multi-inference demo freezes after long duration run
Known Issues¶
ADASVISION-5643 - Mosaic hangs in edge ai apps multi-input multi-inference when run with both instances of MSC
ADASVISION-5631 - tiovx mosaic doesn’t work if the pixel ratio of input videos is different
ADASVISION-5447 - Models failing due to tiovxdlcolorconvert limitation
ADASVISION-5420 - No valid frames found before end of stream Image Input -> Video File Write
Model Zoo¶
About 83 models from 8.4 model zoo are regressed with multiple input sources such as camera and video files. A list of unsupported models are as below,
- Height/2 is odd, tiovxdlpreproc block fetch logic fails
TFL-CL-0100-efficientNet-edgeTPU-m
TFL-CL-0140-efficientNet-lite4
TFL-CL-0170-efficientNet-lite1
TFL-CL-0190-efficientNet-edgeTPU-l
- High latency causes DSS underflow
ONR-SS-8730-deeplabv3-mobv3-lite-large-cocoseg21-512x512
- Odd resolution, tiovxdlcolorconvert fails
TVM-CL-3430-gluoncv-mxnet-xception
TFL-CL-0040-InceptionNetV3
- Human pose model with unsupported post processing
ONR-KD-7000-human-pose-ae-mobv2-fpn-spp-udp-512x512
ONR-KD-7010-human-pose-ae-res50v2-fpn-spp-udp-512x512
ONR-KD-7020-human-pose-ae-mobv2-pan-spp-udp-512x512
ONR-KD-7030-human-pose-ae-res50v2-pan-spp-udp-512x512
Note
Please also refer to edgeai-gst-plugins issues for more details on bugs filed on custom GStreamer plugins
Software components¶
List of software components used in this version
Componenet |
Version |
|---|---|
PSDKLA (filesystem, uboot) |
08.04.00.11 |
PSDKRA (firmware) |
08.04.00.06 |
Docker |
19.03.8-ce |
Ubuntu (Docker) |
20.04.LTS |
Python |
3.8 |
OpenCV |
4.5.1 |
GStreamer |
1.16.2 |
Cmake |
3.16.3 |
Ninja |
1.10.0 |
Meson |
0.57.2 |
Jupyterlab |
3.0.14 |
NeoAI - DLR |
1.8.0 |
Tensorflow |
TIDL_PSDK_8.4 |
TFLite-runtime |
TIDL_PSDK_8.4 |
ONNX-runtime |
TIDL_PSDK_8.4 |
PyYAML |
5.4.1 |
TI Model Zoo |
8.4.0 |
edgeai-gst-plugins |
0.7.1 |
edgeai-tiovx-modules |
8.4.1 |
edgeai-tidl-tools |
8.4.0.2 |