Yolov3: Question about image size

[lsd1994]

I want to train my custom dataset with resolution 2880×540.
So how can I change the image size to this? After reading this (#456) , maybe I should add --img_size 2880 in training command line, so the letter box may automatically resize it to 2880×544. Is that right?
Thank you for help.

[glenn-jocher]

@lsd1994 test.py and detect.py automatically use rectangular inference. train.py will use rectangular training if you pass it a --rect argument, i.e.:

train.py --img-size 2880 --rect ...

[glenn-jocher]

@lsd1994 otherwise you'll be training at 2880x2880 if you simply do train.py --img-size 2880

[lsd1994]

I see, thank you very much!

[joel5638]

@glenn-jocher Ive resized the images from 640x512 to 160x120. DO i have to change the annotations? i mean re annotate?

[glenn-jocher]

@joel5638 the annotations are all in normalized coordinates (0-1). See https://github.com/ultralytics/yolov3/wiki/Train-Custom-Data#train-on-custom-data

[joel5638]

@glenn-jocher yeah. But you’re saying that when we resize images we dont have to change the annotations, isnt it?

Just use the ‘—multiscale 160’flag at the training?

[glenn-jocher]

@joel5638 theres nothing to relabel since all labels are in normalized coordinates.

—multiscale does not take any arguments, you simply pass it by itself:

python3 train.py —multiscale

[joel5638]

@glenn-jocher Im trying to train 160x120 jpeg images with ‘—multi-scale’ flag and batch size of 8.

So while training. On the epochs. The image sizes are changing.

Is it the right way or am i missing something?
I dint make any changes in the cfg file too. I let the width and height be 640x512.

Is the model training on right sizes?

So its happening because of the multiscale flag.

What does multiscale does?
Cant we just train 160x120 images and not give multiscale flag?

[glenn-jocher]

@joel5638 of course the image sizes are changing, you requested multi-scale. It says right there in your terminal output using multi-scale 288-640.

[joel5638]

Okay.
So if i want to train with 160x120. I dont have to mention multiscale.
It automatically takes the sizes available in the image folders

[glenn-jocher]

@glenn-jocher training and inference always operate at --img-size. See the argparser arguments:
https://github.com/ultralytics/yolov3/blob/65eeb1bae5ea1f7d7249c0caf581e95de0dc1637/train.py#L398

[joel5638]

Okay.
So for 160x120. that would be
Python3 train.py —data/custom.data —weights/ultralytics49.pt —cfg/yolov3-spp-r.cfg —img-size 160

[glenn-jocher]

@joel5638 yeah! But I would use this instead, since 49 is a bit out of date. You might want to git pull to get the latest also.

python3 train.py --data custom.data --weights yolov3-spp-ultralytics.pt --cfg yolov3-spp-r.cfg --img 160

[joel5638]

Something doesnt seems to work.
Ive stopped the training and checked the inference.

The box doesnt seem to fit in the object.
Trying to understand what is going wrong.
wrong label too.

[glenn-jocher]

Theres no point in looking at inference results before training is complete. You need to look instead at the images mentioned in the tutorials to see if your labels are correct, and if your losses are good.
https://github.com/ultralytics/yolov3/wiki/Train-Custom-Data