[Questions] Looking for some informations about object detection

[elementdl]

Hi,
I'm an engineer student in Electronics, and for the intership of my last year I'm working on autonomous Vehicles, especially on object detection.

My goal is to use real-time object detection on a prototype (a size of the remote car toy (30cm), not a real car) to make him partially autonomous. I will use a Jetson Nano from Nvidia
For now I working on YOLO, and I have some questions :

There is many many algorithms for object detection and sometimes I struggle to see the differences between them.
For now, what is the better algorithm to use in my case? (real-time detection, prototype of the size of a remote car toy (30cm), using jetson nano) I'm thinking about yolov3-tiny. Will Yolov3 will be too "heavy" in order to have real-time detection?

About the many algorithms, in this project what is yolov3-spp, yolov3-tiny-prn and yolov3-openimages?

Would you recommand any dataset besides COCO?

Is it possible to change the output resolution of an algorithm? I suppose I can change the ouput resolution of yolov3-tiny but I don't know how, I suppose It's related to the number of the convolutionnal layers?
Anyway knowing that the prototype will be way smaller than a car maybe a resolution of 416*416 is enough?

My goal is to use object detection to control the prototype, so I have to use the results of the algorithm of object detection to control the motors of the prototype. Is it possible with the framework darknet? Or do I have to change the framework in order to be able to modify the code more easily?

Thanks by advance for your response.

[AlexeyAB]

• Jetson Nano - 0.5 TFlops
• GTX 1080Ti - 10 TFlops

So divide the FPS by 20.

I would suggest you to use Intel Myriad X Neurochip 2 Watt with enet-coco.cfg (EfficientNetB0-Yolov3) model: https://github.com/AlexeyAB/darknet#pre-trained-models

Comparison of different models on Intel Myriad X Neurochip: #5079

Comparison of different models on GTX 1080Ti:

[elementdl]

Hi @AlexeyAB ,

This comparison of different models are based on algorithm using how much classes?
I would like to know how the number of classes would impact the performance of my algorithm.
If I use a custom Yolov3-tiny on a jetson nano with 10 classes maximum, how much would it impact the FPS?

Sincerely,

[AlexeyAB]

Speed of network doesnt depend on number of classes.
Accuracy depends on number of classes.
All these models are trained on MS COCO dataset with 80 classes.