Image Preprocessing for Training and Testing

[nseidl]

Where can I find in the source preprocessing for images in training?
Are images scaled to 416x416 maintaining aspect ratio then padded with gray 128 to make up missing space, in training?

What is the Python Pillow / opencv implementation of resize_image that is used in test_detector?

[primepake]

you can see it in src/detector.c

[nseidl]

First question:
I can see that when running detector test ..., images by default (without -letter_box) are preprocessed using the resize_image function.

darknet/src/detector.c

Line 1322 in aeb9da6

if(letter_box) sized = letterbox_image(im, net.w, net.h);

darknet/src/image.c

Line 1276 in aeb9da6

image resize_image(image im, int w, int h)

What is a similar python (opencv or pillow) implementation of this resize_image methodology?

Second question:
During training, what preprocessing is applied? Is letterbox_image or resize_image used? Where in the source can I see how images are preprocessed during training? I've looked over train_detector several times but can't find where images are preprocessed during training.

[faybak]

hello @nseidl did obtain an answer to your question please?

[nseidl]

hello @nseidl did obtain an answer to your question please?

#3119 (comment)

By default in this repo, images are stretched in training and in inference.

Python (pillow) code:

from PIL import Image
image = Image.open(image_path)
image = image.resize((416, 416), Image.BICUBIC).convert('RGB)
return image