Same FPS for tow resolutions using YoloV3
Hi all, I already posted this question in Nvidia Forum but I didn't get an answer. I am using the Jetson AGX Xavier board with Ubuntu 18.04. Also, I am using YoloV3 with an ethernet camera and ROS to classify obstacles (darknet_ROS). I have a question about the number of FPS. In fact, I change the resolution of the input image, 1920x1080, and 640x480, and I get the same fps, about 10 FPS for both of them. I want to know why I get the same FPS for the two resolutions? I expected more FPS for the 640x480 resolution.
The original post is here: https://forums.developer.nvidia.com/t...
Thank you
Just to confirm, what matters is the image shape for which your Yolo model was trained to (in the config file at the time of training) and not size of the image you actually input. When used, YoloV3 will first resize the input image so that it matches the shape it was trained for (or maybe some implementations may require you to provide directly the right shape ?) So if your model was configured and trained for 400x400 images, then 900x600 or 188x188 images will be first resized to 400x400 (note the change of ratio can degrade detection accuracy) and process as 400x400 image, it will always takes the time of a 400x400 image. I observed that the processing time is typically proportional to the number of pixels: retraining a model from 400x400 to 200x200 typically makes the model 4x faster.