GiangHLe
commented
3 years ago As a theory, YOLOv3 is a fully convolution model, so any input size will be suitable. However, the author recommends the factor of 32, to make sure the output is the feature map matched with the grid image.
You have to try to know if 768-input-size is better or not, but usually, the bigger, the better and slower, as you can compare the performance between 320, 416, and 608 in YOLOv3 paper.
hello! I want to change model_image_size in yolo.py to 768,768. Do I need to modify input_shape to 768,768 in train.py at the same time? I tried it. After training the model with 416 and 416 shapes, I can directly modify the shape in yolo to 608 or 768, which can be used directly, so that the accuracy is higher.