Can i train model with Rectangle ratio images?

Support YOLO
1

Hi im trying to find USB camera and Thrmal camera module for my Xavier NX developer Kit.

I realized that those modules usually support 4:3 ratio images.

I’m worring about resizing image which could makes my model dumb.I cant find training options about images ratio.

And also im trying to modify v8’s input 1*3640640 into 1*4*640**640 for my thermal camera data

or making additional backbone which is only for thermal data (in this case I will use C2PSA to

integrate two other tensor - one is from RGB backbone and the other is from Thermal backbone-)

Is this dual-backbone idea possible in ultralytics? or should i give up ur pre-trained weight files and make my own torch model?

2

Image size aspect ratio doesn’t matter as long as you’re consistent between training and inference. Ultralytics performs the resizing. You don’t need to resize.

For multi-channel, Ultralytics has support for it through TIFF files:

3

Thanks for replying

what if thermal and RGB has different resolution??

As an example, RGB is 960 * 720 and IR is 256 * 192
both is 4:3 ratio
Does ultralytics automatically resize this too?

4

No. They should be same resolution with the thermal image as 4th channel. You basically need to combine them to 4 channel images