search

zzh8829 / yolov3-tf2

YoloV3 Implemented in Tensorflow 2.0
MIT License
2.51k stars 908 forks source link

YOLOV3 VS SSD_res50_fpn and Inference of jetson nano #161

Open PythonImageDeveloper opened 4 years ago

PythonImageDeveloper commented 4 years ago

Hi, 1- Is it possible to train yolov3 with only one GPU 1080ti? 2- Is it possible to run trained yolov3 on jetson nano? 3- SSD_res50_fpn-640x640 achieve 30 FPS with 1080 ti, and 1 FPS with jetson nano, Can yolo v3 achieve 30 FPS with 1080 ti with input size 640x640? also achieve above 1 FPS on jetson nano? 4- SSD_res50_fpn-640x640 achieve 35 MAP^1 ON COCO Dataset, Can yolo v3 achieve this MAP^1?

AnaRhisT94 commented 4 years ago

1 - Yes 2 - No ( You need atleast 2.5~2.8 GPU RAM) 3- Yes, No idea. 4 - Maybe, Yolo is a little faster - SSD is has a bit more accuracy.

Side note: Use MobileNet detectors for Jetson Nano. Which are suitable.

PythonImageDeveloper commented 4 years ago

Thanks. I don't understand what's difference between metric MAP and MAP^1. In the Tensorflow object detection model zoo API, all models are reported as MAP^1, but in the official yolo website all models are reported as MAP.

AnaRhisT94 commented 4 years ago

Thanks. I don't understand what's difference between metric MAP and MAP^1. In the Tensorflow object detection model zoo API, all models are reported as MAP^1, but in the official yolo website all models are reported as MAP.

I don't know, I don't think it should matter to you too much. It's also off-topic of this repo.