Xinfer plugin supports custom GStreamer video formats for AI tensor buffers.
vvas_xinfer may show these names in gst-inspect-1.0 output, negotiated
caps, and logs such as inference preferred caps. The selected tensor format
is derived from model-format, tensor layout, and tensor data type. For the
naming convention and mapping table, see Custom AI Tensor Formats.
For insight into the implementation details, refer to the vvas_xinfer source code located at <vvas>/vvas-gst-plugins/sys/infer.
Note
To ensure the smooth operation of multithreaded applications, it is important to create the ML pipeline in such a way that the vvas_xinfer instances are sequentially created instead of concurrently by multiple threads.