I already trained a deep neural network called YOLO (You Only Look Once) with high-quality images (1920 by 1080 pixels) for a detection task. The result for mAP and IOU were 93% and 89% respectively.

I wanted to decrease the quality of my training data set using some available filters, then I use those low-quality images along with high-quality images to train the network again.

Does this method increase the accuracy (or, in general, performance) of the deep neural network (for a detection task)? Like mAP and IOU?

My data set is vehicle images.

mAP: mean average precision

IOU: intersection over union ( or overlap)

  • $\begingroup$ Hi and welcome to AI SE! I don't know if it will improve performance, but data augmentation (e.g. the addition of Gaussian noise to the images) can potentially make your model more robust. Maybe someone will provide later a more formal answer with more helpful content. $\endgroup$
    – nbro
    Mar 17, 2020 at 22:38


You must log in to answer this question.

Browse other questions tagged .