Which pretrained model to use on Jetson Nano?

[user151765]

Hi,
I am new to Jetson Nano. I just wanted to know which pre trained deeplearning model will give me greater than 60fps on 1400x1401 resolution image.

Kindly reply.

[meeki007]

none that I know of because of your 1400x1401 resolution image.
ya down scale it to 640x640 and use yolov5n ya might get 60fps… not likely. Depends on the model you train for it and how many objects your trying to detect in each image.

Start reading … allot… dive down the rabbit holes. try some code.

Model	size	objects	mAP	Jetson Nano 2015 MHz
NanoDet	320x320	80	20.6	28.2 FPS
YoloFastestV2	352x352	80	24.1	38.4 FPS
YoloV2	416x416	20	19.2	10.1 FPS
YoloV3	352x352 tiny	20	16.6	17.7 FPS
YoloV4	416x416 tiny	80	21.7	11.2 FPS
YoloV4	608x608 full	80	45.3	0.7 FPS
YoloV5	640x640 small	80	22.5	4.0 FPS
YoloX	416x416 nano	80	25.8	17.6 FPS
YoloX	416x416 tiny	80	32.8	8.3 FPS
YoloX	640x640 small	80	40.5	2.6 FPS

GitHub - Qengineering/YoloV5-ncnn-Jetson-Nano: YoloV5 for Jetson Nano

[AastaLLL]

Hi,

Due to Nano’s limited resources, it’s recommended to try our TAO toolkit for some pruning mechanisms.
You can find the performance of these models on the below page:

NVIDIA TAO Toolkit

Thanks.

[user151765]

I just want to detect only 2 objects for the time being. Yolov5 is a bigger model right? I cannot use tensorflow/tensorRt I am supposed to use only opencv.

[AastaLLL]

Hi,

Maybe you can try some lightweight models, ex. YOLOv3 tiny, to see if it can meet your requirement.
Thanks.