What is the specification of the preset yolov3_tiny and resnet10 inference model in DS 5.1?

Accelerated Computing Intelligent Video Analytics DeepStream SDK
#1

• Jetson Nano with DeepStream 5.1
• JetPack Version 4.5.1

So i did some tests on primary inference, then I look at the inference txt file. The path leads to /opt/nvidia/deepstream/deepstream-5.1/samples/models/Primary_Detector. Here I found resnet file called resnet10.caffemodel with its engine and the label. That’s only on resnet10.
For the yolo part, the path is in /opt/nvidia/deepstream/deepstream-5.1/sources/objectDetector_Yolo/. I found the .weights file yolov3-tiny.weights with custom label.

I assume those are pretrained. Are there specifications for those inference models like the datasets used, train/val/test amount, learning rate, etc.?

#3

Regarding models under objectDetector_Yolo, you could find its introduction in Custom YOLO Model in the DeepStream YOLO App — DeepStream 5.1 Release documentation

#4

I found in objectDetector_Yolo, the preset for YOLOv3-tiny uses pretrained COCO dataset. I guess that’s fine for me, thank you.

As for Resnet10, care to explain how it’s trained? I searched at Performance page on Deepstream documentation, I assume this one in the folder is what’s called DetectNet_V2 -Resnet10, isn’t it?

#5

Sorry for delay! This is trained with NVIDAI internal data.
You could find more info in NVIDIA NGC

Thanks!