Open lfcnassif opened 3 years ago
This will be great.
some references: https://github.com/diovisgood/agender
Dataset for testing: https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/
Dataset for testing: https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/
Actually, most models are trained using this dataset. If I find some with lots of children, I will let you know.
I run some preliminary tests. For the models 4 and 9, results were considered correct if they are within a 3 year error. As for the model 6 the result is considered correct if the age falls into the predict interval. I run these models over a 1000 images of the IMDb data-set and the results were:
Model | Correct | Incorret |
---|---|---|
Model 9 | 460 | 560 |
Model 4 | 229 | 771 |
Model 6 | 162 | 838 |
Well, that is worse than I expected. Could you measure the results using precision x recall? I think we should target high recall values, considering CSAM cases.
Maybe @joaomacedo could give some advice here.
Removing the assignee as this is frozen.
After testing, improving and merging #361, we can store faces bounding boxes and run some age estimation algorithm like https://www.pyimagesearch.com/2020/04/13/opencv-age-detection-with-deep-learning/
The ImageViewer could also be extended to draw faces bounding boxes stored in items and display ages.
This will be very useful for CSAM cases.