Tao input size yolo_V4

[david38]

Hello, I want to train an yolo_V4 on a custom dataset. I made a post TAO augmentation question, but I need more info now.
So if I set in the training config the following:

output_width: 600
output_height: 600 
output_channel: 3

And if I have images of size 320x340, 600x400, 800x800 (for example), how would preprocessing will work? What algorithsm for upscale, downscale will be used? Also, I have bounding boxes, will TAO modify the bounding boxes also?

Thanks!

[Morganh]

The augmentation method is mentioned in YOLOv4 - NVIDIA Docs

@david38

I have bounding boxes, will TAO modify the bounding boxes also?

Yes, the TAO can modify the labels as well.

[david38]

It says nothing about how rescale is made…I read the documentation before wrinting here… but is not complete.

Why is so hard to respond to my questions?..
If I have an image 320x400 how will Tao rescale this image to 600x600.
If I have an image 800x900 how will Tao rescale this image to 600x600. (bbox also)
You understand my question now?.. and this is not about augmentation at all…is about how TAO deals with a dataset with different sizes…

[Morganh]

In short, it is using resizing finally to meet the output_width and output_height. But before resizing, it is really related to augmentation. TAO will check if vertical_flip/horizontal_flip/jitter/mosaic/etc are enable. If one or several of them are enabled, the image and label will do the augmentation accordingly.

[david38]

How is this “resizing” made? what procedures? where in the documentation is made how “resizing” is made?

[Morganh]

There is no update from you for a period, assuming this is not an issue anymore.
Hence we are closing this topic. If need further support, please open a new one.
Thanks

After the images/labels are augmented accordingly, TAO yolov4 will use cv2 to resize to output_width and output_height. This is not mentioned in user guide yet.