1. What is the difference between face detection, face recognition and face classification?
Although related, there is a difference between face detection, face recognition and face classification:
2. I want to try VisageCloud. What do I do?
3. Ok, cool. So what is this face recognition thing good for?
There are many use cases for face detection, classification and recognition:
4. How is VisageCloud different from other face recognition services?
VisageCloud uses state-of-the-art face recognition technology, based on research done in 2015 and 2016. Face recognition services which are older than that probably use older models, such as linear quantifiers and principal component analysis, which are slower and less accurate.
In our tests on large public datasets, such as LFW (Labeled Faces in the Wild), we we able to achieve a 94-96% correct recognition rate.
Moreover, our approach to managing data is a lot more open: calling our API methods will return the actual feature map (think of this a a face signature or a face hash) so you can store it, handle it, write your own logic with it.
5. Do I have to keep my data in the cloud? After all, your service is called VisageCloud.
Unlike other vendors who only let you use a specific cloud providers to run their face recognition service, we encourage you to choose and find the best fit for you.
You can install VisageCloud on-premise running on your dedicated hardware, you can deploy it in your private cloud, on a public cloud provider of your choosing or you can use our deployment, by just requesting an API key.
Ultimately, it depends on your use case and your business needs. If you want to test this whole face recognition thing without making big commitments, you can try our deployment. If you have an application which requires very low latency or if controlling the data on your own infrastructure is important, go for a private deployment.
Whatever your priorities, we encourage you to contact us and we’ll work side-by-side to find the solution architecture which suits you best.
6. Does VisageCloud API store my pictures?
The short answer is “no”.
The algorithms employed by VisageCloud do NOT require storing the picture for future searches and recognitions. This works by doing the analysis on the picture, finding the faces and then extracting a feature map for each face. Think of the feature map as a one-way hash function purpose-built for faces, which does not allow the reconstruction of the original face. Thus, we do not need to store the original picture of Joe/John in order to later perform a face search. Our API however allows you to set a parameters to store the original picture as well, but this is done only if you explicitly choose to. Otherwise, the image is discarded from memory after the signature is extracted.
7. Do VisageCloud demo apps store my pictures?
Our demo apps (like the one that allows you to find which celebrity you look like most) stores the original picture (stripped of any EXIF data) for up to 30 days. We only do this so we allow you to easily share with your friends the results of the test. The picture(s) is stored under a unique securely random 64 character key which is only given to you once, when you upload your picture. Unless you share the link with anyone, no one will have access to it. In case you have accidentally uploaded a picture which you simply don’t want stored anywhere else, contact us and we’ll take it down in the shortest time possible. All pictures are automatically deleted after 30 days.
8. Feature map, hash functions for faces - what are you talking about?
An image, even a small one, contains a lot of data. If you have a RGB (color) image that’s 500x500, it stores 750.000 individual numbers. Our algorithms will isolate the face region, extract the face signature and that will result in only storing up to 150 numbers which strongly define that person. Not only is this faster and more memory-efficient when we perform a comparison, but it is also very good for data privacy, since there is no way to “rebuild” the original 750.000 values from the 150 values of the signature.
9. Does VisageCloud support lateral (profile) faces?
While we plan to support this feature in the near future, we do not at this point. As a rule of thumb, a depiction of a face is considered lateral (profile) if only one eye is clearly visible (the other one being occluded by the nose).
10. Does VisageCloud support rotated faces?
Our algorithms support alignment and normalization of faces which are rotated to up to 45 degrees.
11. How much does it cost?
There are a lot of variables involved, such as the number of analyses you perform, where you run VisageCloud (your infrastructure or our infrastructure) and the number of profiles you store. The best way is to contact us directly so we can explore together what the best solution for you is.
12. Do you support any other computer vision features?
At this time, our focus is on face recognition. However, we do love new challenges and we are looking to add new features, starting with periocular face recognition (recognition based on the area around the eyes), iris recognition and additional facial features.
13. Why is the face in a picture I uploaded is not being detected?
While we use state-of-the-art detection methods, nothing is beyond error. The most common causes for faces not being recognized include: completely opaque sunglasses, significant facial occlusion (i.e. another object standing in the way), uncommon rotations (i.e. face is upside down) and, in rare cases, extreme lighting or extreme makeup (eg. think Halloween makeup in a dark club). We are interested in any such difficulties you encounter, so do let us know what’s not working as you expected.
14. How accurate is the face recognition system?
There are a lot of factors influencing the accuracy of face recognition, such as lighting, noisy/out-of-focus images, partial occlusions (such as sunglasses) or complete occlusions (veils, masks, face painting). In spite of these effects, in order to maximize accuracy, consider uploading several pictures for each profile (person). As a rule of thumb, you need 3 reference pictures for each profile to get good recognition matches, 5 pictures for each profile to get stronger accuracy and 10 pictures if you want to get the best possible results. Using more than 10 pictures, in our tests, does not produce further relevant improvements. Of course, the algorithm works reasonably even with one reference picture per person, but it will not yield the best results.
15. My eyes are blue, VisageCloud says they’re brown. What gives?
There are situations when there is simply not enough information to detect eye color. This happens when the area depicting one eye has only 1-10 pixels in between. This causes the pixels that define eye color to get blurred with their neighbors and sampled down. Basically, it may just happen that there isn’t one blue pixel in the eye area.
Another potential cause of color space distortions can come from lighting or filters (think Instagram) applied to the image. You cannot accurately describe the difference between green and blue eyes in a black&white or a sepia picture. While our algorithms work to produce the best description of colors, this is limited by the information which is present in the original picture.
15. My hair has several colors. What will VisageCloud see?
VisageCloud can only see one hair color for each face. Should there be more colors present in the picture (blonde with dark roots, ombre, half-pink-half-green), it will choose those colors which are best represented towards the midlines of the head.
16. What should I not upload?
We ask you not to upload pictures of people under the legal age limit in your jurisdiction or who are under 18 years old. You may not upload pictures without the consent of all the people depicted. Please take the time to review our Terms of Service before uploading.
17. Someone uploaded a picture that shouldn’t be online. What can I do?
Contact us and we’ll take it down immediately.
18. I want to build my own face recognition theme park. Where do I start?
VisageCloud is not open source. We are considering making some parts of it open source in the future. In the meanwhile, we can tell you how we got started: by reading the Google Net Inception Paper and by playing with TensorFlow and Torch. You can also check out these curated lists of deep learning and computer vision resources here and here.
Let us explore together how VisageCloud can best work for your use case