Support for Test Time Augmentation with weighted-box fusion

Python
#1

Is there any function in open-cv to undertake Test-time Augmentation along with weighted-box fusion for yolo.?

#2

you mean - something like this or this ?

(then the answer is probably: no, you have to implement it on your own)
((paper mentions only horizontal flipping for augmentation, really, that’s just a flip() away…))

1 Like
#3

Well, there could be multiple augmentations and then the final ensemble-based weighted-box fusion. I tried implementing augmentations followed by weighted-box fusion but not getting accurate results. hence, wanted to confirm it based on the functionality in open-cv.

#4

you mean, an ensemble of original (1st set) and augmented (2nd set) single detector results ?

#5

yes! An open-cv implementation could be faster.

#6

just curious … if you flip an image and detect something, – you have to flip the box coords back, right ?

and i guess, the success of augmentation depends much on how the model was trained. did it have images “from both sides” ? any similar augmentation in the training phase ?

#7

Yes, it will be good to include those augmentations in the training process so that the model performs well on it.