Convert Tensorflow Object Detection Model F-RCNN ResNet50 to TensorRT, is it possible?

Robotics & Edge Computing Jetson & Embedded Systems Jetson Xavier NX
,
1

Hi there,
has anyone tried yet to convert the F-RCNN ResNet50 to TensorRT?
It’s inference with the saved model is incredible slow on pure TF2 Backend/Object Detection API and eats lots of RAM…

Or maybe any other advices for inference optimization on the Jetson NX? The model’s accuracy is just great, so I don’t really want to chose another.

Greetings, Julian

3

Hi,

We have two Faster-RCNN example, please give it a try:

From .pb file : /usr/src/tensorrt/samples/sampleUffFasterRCNN
From caffe file: https://github.com/NVIDIA/TensorRT/tree/master/samples/opensource/sampleFasterRCNN

Here are some benchmark results for XavierNX.
Although there is none from FasterRCNN, it can give you a rough idea about the performance of TensorRT:

Thanks.

4

Hi AastaLLL, thank you for the reply! Are the samples you mentioned in your post comaptible with Tensorflow 2.x (saved model format)?
Are there some samples for Python?