This is a bit of a hack solution has I think the correct solution is to
expose model caps on sinkpad (eventually sinkpads). Till then I think
this is reasonable.
- Add a property to onnxinference to set datatype.
- Fix internal buffer allocation size based on datatype.
- Extract method to remove alphe channel and convert to planar image
when requested. Also template the method to support writing to buffers
of different datatype.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5761>
This element refactors functionality from gstonnxinference element,
namely separating out the ONNX inference from the subsequent analysis.
The new element runs an ONNX model on each video frame, and then
attaches a TensorMeta meta with the output tensor data. This tensor data
will then be consumed by downstream elements such as gstobjectdetector.
At the moment, a provisional TensorMeta is used just in the ONNX
plugin, but in future this will upgraded to a GStreamer API for other
plugins to consume.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/4916>
