Image size

[sanazss]

Hi, Thanks again for your wonderful work. I have a question regarding image size in yolov3. The default in training and testing is 416 and there is augmentation for resizing the image in dataset.py. Are images resized to 416. For example if I have an image size of 256 then the size would be changed to 426 in the network?thanks

[glenn-jocher]

Yes all images are letterboxed (resized and padded) to img-size during training.

[sanazss]

Thank you for your reply. So if I train by img-size 256 all images will be 256 or 416?

[glenn-jocher]

@sanazss python3 train.py --img-size 256 will resize all of your images to 256x256 with padding added to the short side. Aspect ratio is maintained. See #232 (comment)

[sanazss]

Thanks. So if all my images are same size I don’t need to apply the letterbox for padding. Right?

[glenn-jocher]

@sanazss letterboxing is automatically applied if needed. If not needed its not applied.

[sanazss]

Thank you for your answer Glenn. I have another question about the configuration file. I use yolov3-1cls cfg file which has an image size of 416. so does that mean the input image size would be resized to 416 in the network? should change the image size in cfg file?

[glenn-jocher]

@sanazss the training settings in the cfg are not used. cfg is only used to define model architecture. Training settings are defined in train.py.

[glenn-jocher]

@joel5638 python detect.py --img 608

[glenn-jocher]

@joel5638 hmm I don't understand. Are running python on the edge device? Is this iOS or Android?

The normal worflow would be to export to onnx, and then from there to coreml or tflite.