Generate perturbation for arbitrarily sized images about tog HOT 4 CLOSED

Which model are you using? If it is YOLOv3 or SSD, resizing inputs is necessary because these techniques require a specific resolution of the input image (e.g., for SSD300, it must be 300x300). If you are using Faster R-CNN instead, you can modify the image preprocessing component and send the image with the original resolution to the model.

from tog.

I'm using the YOLOv3 model and attacking a YOLOv3 model trained using darknet. The problem I'm facing is that even though the darknet model downsamples images to 416 x 416, it loses most of its performance on the images that have been processed by the letterbox_image function

from tog.

Unfortunately, this downsampling is requested by YOLOv3. There are two solutions you may explore:

1. Use YOLOv3 with a higher resolution (but it should be a multiple of 32)
2. Use Faster R-CNN as it doesn't require downsampling

from tog.

Right! I'll try these options.

from tog.