objectDetector_Yolo in DS 4.0 on Xavier


I did a test about this sample, It was good but I am curious about Implementation to create this source by using IPlugin.

because I remember the possibility to parse 3 kinds of format, UFF, ONNX, Caffe in TensorRT.

so I remember TensorRT Yolo example changed to ONNX.

but I think the config file like config_infer_primary_yoloV3.txt in this source used “model-file=yolov3.weights”

I think this means caffe model

how does it work?


The YOLO sample doesn’t use an intermediate model representation.
It create TensorRT layer directly by the one-by-one matching.

In short, it doesn’t use the model parser within TensorRT.
It create the TensorRT layer directly based on the knowledge of YOLO model.