Module doesn't work if BEV RGB images sizes are not square?

[chowkamlee81]

I need to change FOV from [0,50] and [-25,25] to [0,40] and [-40,40]. In trhis case iam assuming BEV RGB images are of size 480*960, then YOLO modules crashes. KIndly let us know whether YOLOv3 works if images are not of shape square.

[chowkamlee81]

@ghimiredhikura , can you please suggest

[abhigoku10]

@chowkamlee81 did u re-train the model after changing the FOV ?

[chowkamlee81]

@abhigoku10

I tried to retrain by changing the FOV but module got crashed since Anchor gride size is of squared size. Is there is any possibility it will work if we change our resolution to not to a squared size?

[abhigoku10]

@chowkamlee81 ideally the complex yolo takes in non squared resolution, i think we need to generate new anchor boxes for your custom dataset . Can you share the error which your getting

[ghimiredhikura]

Hi @abhigoku10, @chowkamlee81,

Yes currently the model only supports square size, as size is passed as single value (not width and height separately) and used as same value across x and y axis. If you want to make it work with non square size you have to pass height and width separately in models.py file and change the code accordingly.

PS: There was small error regarding input image size in train.py, fixed with last commit.

Best, Deepak