Image detector using non existing images

I am making some custom image detector for a project.

I labeled the images using Matlab imageLabeler, but after the labels were generated we changed the path of the images and moved some of them to another folder for testing purpouses.

For the traiing of the detectors I generated a table with the corrected file names and the labels (similar to the yolo example) and I tought that the missing images will be ignored (they do not exist in the specified path, nor have I included its path in any way). The thing is that training with the complete set (existing and non existing images paths) yields better results than removing the non-existing file names.

Any idea why this is happening?

Answer 1

[Yogesh Khurana]

I think when you are passing full directory as data for training and then you are testing for a part of same data. The accuracy will be greater as you are considering the same data for testing again. The model already trained itself for the data that you are using for testing. Please do not passing the test dataset for accurate results of your code.

Please refer to Yolo example once again for path that you need to include while training and testing.

Hope this helps!