Readit News logoReadit News

Deleted Comment

jedimastert · 6 years ago
For the actual removal, you could try using conent-aware resizing[1] along with blob detection to give the person to be removed a negative weight.

[1]: https://www.youtube.com/watch?v=qadw0BRKeMk&t=3m43s

dsr_ · 6 years ago
I wonder how well it works on people who have a skin color closer to their hair color.
swish_bob · 6 years ago
I was about to suggest it really needs to at least be tried on people with darker skin tones.
Double_a_92 · 6 years ago
Or old white people.
ovi256 · 6 years ago
Yeah, his baldness detector is not robust as it just checks the average color of the area above the forehead, but you could make a robust one through transfer learning a deep model trained on ImageNet further finetuned on a small dataset of a few dozen images of bald and non-bald.

Something like this: https://www.kaggle.com/dansbecker/transfer-learning

This method gets good, robust performance with just a few minutes of training.

benj111 · 6 years ago
I was thinking this is a potentially racist algorithm.

Worked well for Amazon!

https://www.reuters.com/article/us-amazon-com-jobs-automatio...

cyborgx7 · 6 years ago
Since trained classifiers are already being used, and therefore fair game, a trained baldness classifier would be a better choice, I think. Of course you would need a dataset for that.
mkagenius · 6 years ago
Deep learning with a 100 samples can do this completely with the body too. This is a similar project I did: https://comic2gif.com where I remove the speech written on the images, the same technology can be used here.
nacs · 6 years ago
But edge detection is a lot easier on comics than on photos I would think. The background on a comic would also have a mostly solid color background which makes it easier.
mkagenius · 6 years ago
> But edge detection is a lot easier on comics than on photos I would think

That should be true. But deep learning based models work quite good on natural images as Vgg dataset is mostly natural. This is perhaps one-off usage/application on comics.

cantcomplain · 6 years ago
I know this was a joke but I still wanted to see more examples
_cr1k · 6 years ago
hey, author here, thanks for the attention on this post, especially the feedback regarding CV techniques. i wanted to write about some basic computer vision tools in the context of a Seinfeld reference. it was intended as a joke, but i realize it could be hurtful. i'm sorry if i hurt anybody's feelings.
smashingfiasco · 6 years ago
I really like that you explained the algorithm and decisions you were making instead of just ploping code some on the page.

Well done!

Grustaf · 6 years ago
As a photo store clerk I can confirm that this is indeed one of the most frequent requests we get, so this could be a huge timesaver. But you really need to work on the removal part, it’s practically unusable as it is!