EdgeFirst Perception for GStreamer v0.3.0

Downloads

Note: Binaries for this library are built as part of the
meta-edgefirst Yocto layer.

NV12 plane offset corruption — edgefirstoverlay now uses GstVideoMeta
(authoritative buffer layout) instead of GstVideoInfo (tight-packed assumption)
for plane strides and offsets. Fixes a ~10-pixel magenta band at the top of the
output when decoding H.264 via v4l2h264dec on i.MX 8M Plus (Y plane height
padded to 1088 for macroblock alignment).
Split-box tensor detection — the min-dimension heuristic now handles both
TFLite features-first ([4, 8400]) and Ara-2 anchors-first ([8400, 4]) tensor
shapes, enabling auto-config for YOLOv8 segmentation across both frameworks.
Box quantization scale — the pixel-space scale ÷ model_input_size adjustment
is now gated on normalized=false. TFLite uint8 models that already quantize
boxes in [0, 1] are no longer collapsed to sub-pixel slivers. Ara-2 DVM pipelines
must set normalized=false until future DVM exports adopt normalized coordinates.

Auto-letterbox — when the letterbox property is unset, edgefirstoverlay
computes an aspect-preserving fit from the input video dimensions and the inferred
model input size, matching what edgefirstcameraadaptor letterbox=true produces.
The explicit letterbox property still overrides when set.

TFLite NHWC protos produce vertical-stripe mask artifacts on i.MX 8M Plus due to
a layout mismatch between the GStreamer overlay's NCHW shape labeling and the
physical NHWC memory order. Detection boxes and auto-letterbox work correctly;
only mask rendering is affected. Requires a HAL-side fix to handle NHWC protos.

Full Documentation: https://github.com/EdgeFirstAI/gstreamer#readme