Multi-Camera Inference Pipeline — four-camera camera-source inference workflow. See Four-Camera MIPI Streaming Use Case for details.
vvas_xtilecompositor — zero-copy DMA-BUF coordination that allows up to four BGR/RGB camera sources to write into one shared display buffer without copying the source frames.
Asynchronous VART Inference — queues the next frame’s NPU inference while the current frame’s inference is still running, keeping the NPU continuously busy for better performance. See vvas_xinfer configuration.
Segmentation Inference — VART backend now supports segmentation in addition to classification and detection.
CPU Subgraph Support — models with CPU subgraphs (in a single rai file) are supported transparently; no additional configuration/settings change required in VVAS.
Unified GStreamer Logging — vvas-core and vart-x logs now route through
GST_DEBUG. See VVAS Debug Support.