Over the last year or so, I've gotten a decent bit of feedback across a number of different channels. I thought I'd share that here so folks can get a sense of what I'm hearing. This blog post is long, but there are three main categories of feedback covered here if you'd like to skip around: current user feedback, reviews, and exit poll feedback. If you wrote in, please check it out - I've likely responded to your input. I'd like to thank everyone for their feedback, as it has been generally constructive and helpful!
What would current users like to see?
There's a feedback link inside the addon - I'll list key items of feedback (with perhaps light editing) and respond here.
Loving the extension so far, but I'd like to ask if it's possible to add a toggle to quickly enable and disable the extension, either on a per-page basis or globally. While the "trusted" toggle seems to work fine, sometimes I'd prefer to temporarily disable the extension without going through firefox's menu. Thanks.
-someuser
Good news - there is a feature for that! If you go to Options -> Show an on/off switch from the main menu, it will add the following checkbox:
You can then turn the plugin on and off from there. I don't expose it by default because it's not an option you want kids accessing by default. Note: if you turn the setting on or off, you will need to refresh the page while holding the shift key in order to see a specific image blocked/unblocked.
In Trusted mode this image is hidden https://photos1.similarworlds.com/00/00/00/00/03/13/06/57/CdAshley-OCIh4VdpGDPFGRf.jpg In Untrusted it is shown For this image https://photos1.similarworlds.com/00/00/00/00/03/13/06/58/CdAshley-k7spRy6Ix3WLnC1.jpg the opposite happens In my opinion neither should be hidden at all but certainly not in Trusted mode.
-Nina Lanyon
Similar to the note about the on/off switch, you have to do a refresh while holding the shift key in order to see the changes take effect. Note that when I test it with those images here, it's as you would expect: the first image blocks in both Trusted and Untrusted, and the second image shows in Trusted and blocks in Untrusted. I don't think I've fixed any bugs around that behavior, but perhaps? Give it a try and let me know if you're still seeing it Nina.
It's also a good point to bring up about the subject matter. One could make a good argument that neither of these should be blocked; however, let's take the first as an example as it blocks even in Trusted. To a certain degree I can only guess what the AI model is doing, but here's what I think: there are a number of aspects this photo has that match an amateur NSFW photo - the setting, the head not being included in the photo, the pose. My hunch is that the same photo in a professionally shot environment with just the waist down may have not flagged.
Non-human porn match rate is very low. Now I realise that this is a niche topic, and it's unlikely to crop up in most searches + the style and shape of the images is likely significantly different enough that covering this would require an entire dedicated effort to train just for it. But, for instance, on a page like [redacted, description of cartoonized NSFW images slipping through about half the time]. Like I said I don't necessarily expect you to support this, and apologies also if you're already aware of it, but in case you weren't I thought I'd give a heads-up for a potential blind spot in this otherwise very interesting extension.
-Chuck
This is an interesting one, because the dataset for the model actually does have quite a few of this sort of thing in it. While more could be added, I also think there is another contributing factor here: the cartoon nature of the pictures. While the human eye views cartoons as arguably simpler and even easier to recognize than photos, they are by nature more abstract - and harder for convolutional neural networks to fully interpret the same way that we do. As such, some of the anime detection also suffers - the model is just not as good at recognizing these.
hello, I would like to start by saying that this is without a doubt one of the best extensions I have ever seen in my entire life. I have an addiction to pornography and it helps me a lot when I go to a site like 4chan where there are [redacted, tempting] threads. I would like to ask a few things if possible to improve even more in my opinion. 1. block with click, sometimes it doesn't work and if possible I would like to click on the image with the right button and be able to block or allow. (very important) 2. Filter by site, basically I can choose which sites the extension would work with. 3. change the silent mode images, it would be my preference for me to choose my own images for this mode. there's nothing wrong with the images just that i would like to have the ones i prefer (ex: cars) thank you for your work and i wish you success and sorry for my english.
-qvim
Qvim, that's some great feedback. Thanks for the kind words and interesting ideas.
For #1, there's a GitHub issue with a basic implementation but it doesn't "remember" the hidden image.
For #2, I've created a GitHub issue as this is a recurring feature request
For #3, this has similarly been requested a couple times and there's a GitHub issue.
In all cases, I'd recommend following up directly on the GitHub issues themselves if possible.
On 4chan and similar websites were images have a low quality compressed preview of the image the AI isn't able to pickup NSFW content most of the time but almost always picks up on the clickable image link from it. My idea would be for the AI to check images that have an embedded image url to them and if those contain NSFW content then block the preview as well.
-Gerbil
This is an interesting feature request. It's been noted by others as well that sometimes the thumbnail is/is not blocked, and then content is different. I'll have to think about this one a bit, but it's not immediately clear to me how this would work out since there are a number of ways for the HTML to format this sort of thing; additionally, the blocking currently works primarily by considering individual requests.
A setting to optionally block small thumbnails or set an arbitrary minimum/maximum would be nice. The thumbnail for the image on this webpage does not get blocked, when you click it to expand, it IS blocked. Your extension is so awesome, I hope you may find time to add this feature in the future, thank you!!!
- Sneed
As a note, the current behavior is
to skip scanning images that are either 1) less than 1024 bytes or 2) less than
36 pixels on both sides - after some experimentation, this seemed like a sensible default. There's a GitHub issue out there to continue the discussion.
Hello there Zephyr, It's pakxo again. I have an issue while using Mattermost. The extension becomes so so slow and sometimes break, often times break on Mattermost. Even tho, there's 0 NSFW media. It breaks more on my PC tho. It runs Windows 11, 4790K and a 1050ti and sometimes on my Matebook X, I5-10210U with integrated graphics. I don't think it's a performance issue. It's more like a bug or something. Thank you for working so hard, making and maintaining this extension!
-pakxo
I don't use Mattermost but gave this a shot to try to reproduce it. Unfortunately I couldn't get it to break, but if I can get some more feedback over at the GitHub issue I'll take another look!
It's a bit sensitive with YouTube videos, and when it 'detects' imagery (when there is none), the video buffers and will not stop, no matter how much I refresh it, etc. This happens with every second YouTube video or so. Otherwise this is a really helpful little thing and thank you for developing it.
-poly
Video blocking being enabled has very poor performance when open the discord "GIF" tab and browsing GIFs. Disabling video blocking seems to resolve the issue but doesn't block GIFs obviously. I assume this is because there are a lot of GIFs to check at once but thought I'd report it. Great extension by the way, thank you.
-anonymous
Video filtering isn't as good as image filtering yet - but I have it so that you can turn it off separately in the options if it's causing any performance issues or you just don't find the filtering quality to be good enough.
This is really good stuff. May Allah guide you to the right path :)
-knarkzel
Thanks, much appreciated!
I don't think it works with webp images? When I try to load github.com some images don't load. Same with zillow
-Steve
After this feedback, I tried poking around on webp images and wasn't seeing any issues; I also use GitHub and Zillow frequently and hadn't run into anything. I believe there might very well be a problem but I need help finding it - if anybody can post a specific URL or page with a webp image in it that is having an issue - please let me know!
Can you ignore svg images?
-JonNot currently.
Reviews on Addons.Mozilla.org
Overall, the ratings are like this:
Generally users enjoy the addon and find it helpful, however....
Back in November, I shared the addon on HN - the community enjoys seeing others' side projects, and it can be a way to help spread the word. Some folks did try the addon, but during this time it got the only 3 ratings less than 4 stars - a bit tougher crowd! And the discussion on HN was a bit less helpful than usual, with most comments tending to focus on the usefulness or value of the addon's goal rather than technical aspects or new use cases.
However, I'd like to thank Drago here recently for the kind review and the bug noted in the review - the recent release 3.3.2 was a direct result of this.
When new users don't stay, why are they leaving?
When users try out the addon and then uninstall it, an exit poll is opened. This is particularly important because it helps me understand what is driving away new users. Here are the top results:
- 11 votes - Not sensitive enough in blocking sexualized content
- 8 votes - Blocks too many safe images
- 6 votes - Scanning speed
- 6 votes - Missing features
- 5 write-in comments - Complaints about the hidden tab
- 3 votes - Lack of way to give feedback about predictions
So we can see that overall model accuracy - for both reducing false negatives and to a slightly lesser degree false positives - remains a critical concern. Scanning speed is a close second. These top three items - false negatives, false positives, and scanning speed - are all largely based on the quality of the input data, the architecture of the neural network, and efficiency of the scanning process. I've worked hard to optimize all three, but the feedback here lets me know this is a continued area of interest. Unfortunately, general machine learning research has slowed regarding mobile-based image recognition - perhaps because it is largely viewed as solved for ImageNet.
Among those listing missing features as the reason for leaving, a common theme was around manually blacklisting and whitelisting sites and images.
A somewhat unexpected result was how frustrating the hidden tab is for users. For those who have been using the addon for quite a while now, you may remember that this did not exist prior to Firefox 83 - unfortunately it needed to be added because a bad interaction between Firefox and Tensorflow.js caused the addon to basically become unusable. Annoyance is expected, but to see this be the key reason that approximately 16% of those responding say that they are leaving is surprising. Based on the responses, I also believe some users may either be closing the tab without reading the warning or perhaps experiencing some crashes. Some of the feedback was as I'd expect:
Hidden tab nag is annoying
Other feedback was ... a bit more colorful:
the f******* hidden tabs notification kept popping up every 10 seconds, why tf is the tab restarting every 10 seconds
Obviously, this is something causing a lot of frustration. If you are experiencing issues with the tab, please let me know in a comment here or if you are able, file a bug on GitHub.
Conclusion
Once again, thanks for using the addon and providing helpful feedback! I look forward to hearing from you all once again in 2023.
No comments:
Post a Comment