Resnet-10 based SSD architecture

Hi, I was wondering what kind of architecture was used to create the resnet10-ssd that is used on the DeepStream examples.

And please don’t point me to the Transfer Learning Toolkit, because it’s on Early Access and I haven’t got a confirmation yet.

A simple: “It’s based on Deconvolutional Single-Shot Detector” or “You can check out the layers on this file” is enough, but please point me out in the right direction.

Thanks!

Hi,

Resnet-10 is a caffe-based model. You can check the model architecture directly on this prototxt:
/opt/nvidia/deepstream/deepstream-4.0/samples/models/Primary_Detector/resnet10.prototxt

I don’t think the model is based on the SSD.
It’s much close to the idea of DetectNet, which outputs the bounding box directly.
https://devblogs.nvidia.com/detectnet-deep-neural-network-object-detection-digits/

Thanks.

Thanks for the info!

I assumed it was a resnet-10 based ssd because of the benchmarks presented in https://developer.nvidia.com/embedded/jetson-nano-dl-inference-benchmarks

I’ll take a look at DetectNet. Thanks again!

Hi,

The model in the benchmark is the standard resnet(feature-extractor)-SSD.
Different from the resnet-10 in Deepstream SDK.

Thanks.