Top 5 things to know about machine learning content moderation
If you are in charge of moderating user generated content, whether video, image or text based, you may be wondering how you can take advantage of machine learning without having to completely rearchitecting your platform. This can be done by plugging into the API-based machine learning content moderation provided by AWS. Here are the top 5 things you should know about it:
- Video, image, & audio moderation is done by Rekognition
- Machine learning content moderation costs around $0.10 per minute
- There are 10 major categories, and 25 secondary categories AWS tracks
- You can tailor the content moderation to fit your parameters
- You can use the same algorithm for metadata tagging
1. Video, image, & audio moderation is done by Rekognition
In the AWS machine learning space, the workhorse of the content moderation is Rekognition. Rekognition is a deep learning algorithm powered and trained by millions of hours of video, and countless numbers of images which can do robust content moderation out of the box. The system is available on demand (see pricing below) via an API which can be setup by creating an AWS account and providing the right IAM user permissions. If you need help setting this up, please don't hesitate to reach out.
2. Machine learning content moderation costs around $0.10 per minute
While visual machine learning can be complicated to setup, the infrastructure costs on Amazon are often negligible. For example:
- image analysis generally ranges between $0.001 per image (for the first million images) up to $0.00025 per image if you are processing 35 million images or more.
- Basic label detection starts at $0.0010 per image, aka.1,000 images for 1 dollar.
- Video analysis is generally prices per minute per service, with all of the following services billed at $0.10/min: label detection, content moderation, text detection, face detection, celebrity recognition, face search, person pathing
More robust pricing can be found on the Amazon Rekognition pricing page, but even if the calculations get complicated, the price is still incredibly low. As an AWS Reseller, Metal Toad can likely get you better pricing as well. Feel free to reach out and see what kind of pricing we can offer!
3. Amazon tracks are 10 major and 25 secondary categories
The Rekognition API will automatically tag content reviewed with the following major or top-level categories (explicit nudity, violence, etc) and (if applicable) the second-level category:
- Explicit Nudity
- Graphic Male Nudity
- Graphic Female Nudity
- Sexual Activity
- Illustrated Explicit Nudity
- Adult Toys
- Female Swimwear Or Underwear
- Male Swimwear Or Underwear
- Partial Nudity
- Barechested Male
- Revealing Clothes
- Sexual Situations
- Graphic Violence Or Gore
- Physical Violence
- Weapon Violence
- Self Injury
- Visually Disturbing
- Emaciated Bodies
- Air Crash
- Explosions And Blasts
- Rude Gestures
- Middle Finger
- Drug Products
- Drug Use
- Drug Paraphernalia
- Tobacco Products
- Alcoholic Beverages
- Hate Symbols
- Nazi Party
- White Supremacy
4. You can tailor the content moderation to fit your parameters
As a content moderator you may find that some content you want to remove or flag for human review, while other content you may want to allow through or simply delete without review. All of this can be setup and customized in the application layer, which means AWS machine learning is extremely flexible: it describes the type of content and the time stamp where it occurs, and you decide how you want to deal with it.
This also means that every application of machine learning content moderation must be extended and customized. If this is something you are exploring feel free to reach out and we can help you scope out the project.
If you need a comprehensive solution that can provide video upload, streaming, and moderation you can read about how to pull it all together using Rekognition + MUX here.
5. You can use the same algorithm for metadata tagging
Amazon Rekognition does much more than content moderation. Although each service requires it's own configuration and incurs additional costs, Rekognition can also do the following "out-of-the-box":
- Facial recognition - this often the first thing people think of when they think of machines doing image analysis.
- Celebrity recognition - in addition to identifying faces, the algorithm knows "celebrities" out of the box. Jeff Bezos or Werner Vogels are perfect examples.
- Sentiment analysis - "that's a face, are the people happy or sad", etc?
- Text detection - the machine can read things like menus, t-shirts, etc.
- Personal Protective Equipment (PPE) - Rekognition can detect the usage of PPE
- Custom Labeling - one of the most powerful features of Rekognition, the machine can label images with things it observes like:
- Activities (running, playing, laughing, etc.)
- Objects (trees, streets, cars, people, etc.)
- Scenes (beach, sunset, city, etc.)
- Concepts (outdoors, fun, etc.)