Hello, I have just gotten started with yolov5 and I noticed that the val_batch*_labels.jpg images in /runs/train/exp* has annotations that are wrong (I used roboflow to annotate if that makes any difference). Here is one of the sections with wrong annotations (there are many more, some of the errors include: incorrect sizes, placement, etc):
Information that may or may not be needed:
Windows 10
300 epochs
s6 model
The train_batch*.jpg images seem to be better (maybe 65% of them are correct, but the others are experiencing the same issues), but I can’t really tell because of how jumbled together everything is.
Did something get messed up when I exported the annotations? Also, is this purely a graphical error or do I have to worry about it affecting my model?
If this is just an error with displaying the bounding boxes, is there anyway to fix it?
Also, this isnt related but, what is generally the better choice?
Use a bigger model (Ex x6), but train for less time, or
Use a smaller model (Ex s6), but train a lot more?
Sorry if I didn’t explain my question very well.
I know that the ‘\runs\train\exp*\val_batch*_pred.jpg’ images are the predictions that the model gave out,
But i’m confused about why the ‘\runs\train\exp*\val_batch*_labels.jpg’ bounding boxes are in the wrong places.
Another example: (taken from ‘\runs\train\exp45\val_batch0_labels.jpg’)
@Volatility you should raise an issue directly with Roboflow support if your val_batch*.jpg images are incorrect, as we don’t control their datasets or exports.
An example of a correctly labelled val_batch0.jpg for COCO128 is here:
@Volatility did you find the fix? I am running through the same problem, and got them labelled using roboflow…my val_labels are off. but on roboflow they seem right…
