What is the actual prediction range of 'vit' pretrained model?

pidahbus / deep-image-orientation-angle-detection

MIT License

82 stars 20 forks source link

What is the actual prediction range of 'vit' pretrained model? #8

Open HaiderAbasi opened 1 year ago

HaiderAbasi commented 1 year ago

Houston, We have a problem.

Yes, I have tested on the test image and I got the correct prediction of 115 degrees.
Testing on a video (frame by frame) I get weird values even less than -180 i.e (-182, -200), etc and also positive values.

Environment I running this in:

Windows 10
Rtx 3050 (Cuda On)
Python 3.8.10 (virtual environment)

(Attaching here a picture of the result of testing on a video)

pidahbus commented 1 year ago

As the model is trained against a regression problem, ideally the prediction range is (-inf, +inf). The print statement predicts the raw model prediction value. You should perform the angular conversion to change the range between [0, 360). For example, negative values such as -182 degrees are valid predictions that are actually (360-182)=178 degrees. Please make sure with respect to the oriented image if the prediction is actually correct or not.