Can we get real-time fps(30fps) with Faster RCNN caffemodel, TensorRT and Jetson TX2?

I’m trying to do a object detection on distant object using Faster RCNN caffemodel in Jetson TX2 board. I heard that Faster-RCNN is accurate on detecting distant objects. I’ve also heard that Jetson-Tx2 and Tensor-RT enables high speed neural computing. So is it possible to get real-time fps or 30fps with these combinations? Please help

Hi,

Suppose you are asking entire flow combining with camera input and display output.
Here is an experiment for Faster-RCNN model and got around 0.9 second per frame.
https://jkjung-avt.github.io/faster-rcnn/

For real-time detection, there are some faster models, ex. YOLO, detectNet.
Thanks.

So That means we cannot get real-time fps(30fps) with Faster-RCNN in jetson-tx2 with tensorRT.

Hi,

Actually, we don’t have concrete answer about this.
We havn’t profile the Faster RCNN model with TensorRT on TX2.

Here are some object detection model we have profiled for your reference:
https://github.com/NVIDIA-AI-IOT/tf_trt_models#object-detection

By the way, we won’t apply object detection every frame but maybe every two or three frame.
This can lower the computational complicity and also decrease the inconsistency from neural network.

Thank.s

Hi,

@AastaLLL you said:
"By the way, we won’t apply object detection every frame but maybe every two or three frame.
This can lower the computational complicity and also decrease the inconsistency from neural network.
"

is it possible to speed up the network (yolo- based) detection on the tx2; is it possible to not use the detection every single frame but only every second, third…? And use a more “simply”/faster approach in the “in-between-frames” like background substraction, segmantic flow, optical flow, or something else, without losing to much on accuracy?

Thanks!
Regards

Hi,

It looks like you file another topic for your new question.
Let’s track the comment #5 directly in the topic 1046242:
https://devtalk.nvidia.com/default/topic/1046242/jetson-tx2/object-detection-on-tx2-speed-up-by-not-using-detection-every-frame-/post/5309668/#5309668

Thanks.