yolotensordec
This element can parse per-buffer inference tensors meta data generated by an upstream inference element
Example launch command:
Test image file, model file and labels file can be found here : https://gitlab.collabora.com/gstreamer/onnx-models
GST_DEBUG=yolotensordec:6
gst-launch-1.0 multifilesrc location=bus.jpg ! decodebin
! videoconvertscale add-borders=1 qos=false ! 'video/x-raw,pixel-aspect-ratio=1/1' !
onnxinference execution-provider=cpu model-file=./yolov8s.onnx input-image-format=chw
input-tensor-offset=0 input-tensor-scale=255.0 qos=false !
yolotensordec class-confidence-threshold=0.1 max-detections=10 qos=false
label-file=COCO_classes.txt !
objectdetectionoverlay object-detection-outline-color=0xFF0000FF draw-labels=true !
videoconvert ! xvimagesink
The original repository of the Yolo is located at https://github.com/ultralytics/ultralytics. For easy experimentation, the models based on Yolo architecture in Onnx format can be found at https://col.la/gstonnxmodels . This model already has tensors name embedded matching default values of tensors-detections-name and tensors-logits-name properties. It's also possible to embed tensor-ids into any model based on Yolo architecture to allow this tensor-decoder to decode tensors.This process is described in the Readme of repository: https://col.la/gstonnxmodels"
Hierarchy
GObject ╰──GInitiallyUnowned ╰──GstObject ╰──GstElement ╰──GstBaseTransform ╰──yolotensordec
Factory details
Authors: – Daniel Morin
Classification: – Tensordecoder/Video
Rank – primary
Plugin – tensordecoders
Package – GStreamer Bad Plug-ins
Pad Templates
Properties
box-confidence-threshold
“box-confidence-threshold” gfloat
Boxes with a location confidence level inferior to this threshold will be excluded
Flags : Read / Write
Default value : 0.4
class-confidence-threshold
“class-confidence-threshold” gfloat
Classes with a confidence level inferior to this threshold will be excluded
Flags : Read / Write
Default value : 0.4
iou-threshold
“iou-threshold” gfloat
Maximum intersection-over-union between bounding boxes to consider them distinct.
Flags : Read / Write
Default value : 0.7
max-detections
“max-detections” guint
Maximum object/masks detections.
Flags : Read / Write
Default value : 100
The results of the search are