Are there other inference models than object detection and classification possible in deepstream, for example some openpose joint detection?
You can try this. It’s caffe model. It’s better wait our DS4.0 release (maybe in 2 weeks) to parse the tensor output.
I already have a suitable model, in the demo programs the DeepStream SDK is only used for object detection and classification. I asked myself, how to create a pipeline to do other types of inference.
Whether I should create a pipeline like [ videosrc -> inference -> fakesink ] and get the inference output from GstMeta data output of the nvinfer element or if there is another solution to do this.
Regarding the former possibility, I was not sure how to write the necessary config file for doing so.
If this is possible in DeepStream 4.0, I will try it with the new version.
The model generates points and scores in openpose, so it should be similiar with object detection. Very excited to wait for 4.0 release :)
Between detection and classification and other inference,the main difference is output parser. TensorRT does all inference in the same way. 4.0 has improvement in gstMeta and open source nvinfer code.