False positives on Coffee cup images

[adnankhan37]

Hello,

We have been using Inception v3 for NSFW detection and we are getting false positives on coffee cup images.

Any suggestion on how can we tackle this?

[GantMan]

WOW! This is a great find. Which model are you using? The default? Does this happen on ALL the models in the dropdown?

[adnankhan37]

I am using Inception v3. Tried the mobilenet v2 also and it seemed to perform well in some cases like this.

For this image, I tested it on Mobilenet v2 and it gave me a higher probability of being porn than inception v3.

I tested the following image also on Mobilenet v2 and it is giving a false positive.

This seems to be happening with all the models in the dropdown.

[GantMan]

This is good to know. I'll add these to the next training set.

[adnankhan37]

Since I am actively using the Inception v3 model in my project, may I know when will I be able to download the updated Inception v3 model trained on these images? :)

Thanks!

[GantMan]

Did you say that one of the other models was better at not being tricked? Because I'm going to add a voting classifier hat will help.

[adnankhan37]

Yes, that would be Mobilenet v2, but just in a few cases, it performed better than Inception v3. But when Mobilenet v2 gave us a False Positive for an image, the probability it gave was higher than the one given by the Inception v3 model.

[adnankhan37]

Hi,

I might have an idea of what is going wrong with the coffee/tea cup images. I found the following repository of the NSFW dataset with girls holding teacups which you have mentioned in the readme file.

Link: https://github.com/EBazarov/nsfw_data_source_urls/tree/master/raw_data/appearance/reddit_sub_TeaGirls

Can that be a possible reason for us getting False Positives on the cup images?

Thanks!

[GantMan]

Ahhhh, that makes sense. Strange! Yeah, if it would be important to counter-balance this kind of data so the AI doesn't form a bias against teacups.

[adnankhan37]

Yes, that sounds great. It would be quite helpful if the model is trained on teacup images and then we can use that updated model.

Thanks!

[adnankhan37]

Hi, is there any update on the training? I found more false positives on the following images. I tested them on Inception v3. Thanks!

[GantMan]

I've got a few other tasks before retraining this model on the new images. My goal will be to create some metrics for retraining. Unfortunately those are behind ANOTHER item I need to complete. If this is a blocker I highly recommend you don't wait for me, and perform transfer learning to fix any bias you find.

Since this is a free open source project, the progress is relegated to my free time for OSS.

[adnankhan37]

Sure, thanks for the help!