trueface/visionbox-gpu repository overview

0.8.4

• adding timestamp of identification for each face to /identify/video endpoint
• bugfix related to Postgres backend integration
• /identify endpoint now takes a video rather than an image. This allows visionbox to finds the largest face in the image and identifies it by its label.
• new endpoint: /v1/detect_glasses which will detect if the person in the image is wearing glasses
• new endpoint: /v1/collection//<collection_id> will list all identities in a collection
• new endpoint: /v1/detect_mask will detect whether the person in the image is wearing a face mask
• feature: /v1/identify/video will now also return the timestamp of a recognized face in the video
• feature: /enroll now takes a flag to store the enrollment image in the database, it's an opt-in feature and can be used to display the image later in a UI
• changed the base image to Ubuntu 20.04 LTS

0.8.3

• bugfix: fix related to Postgres backend integration

0.8.2:

• Added tfsdk_options.json file to docker container. This allows users to set a custom default configuration for their visionbox container.
• sdk face recognition model can now change 'on-the-fly' based on which database the visionbox is accessing. So if you're using a TFV5 facial-recognition model, and then access a collection using the LITE recognition model, the system will switch to the LITE model as necessary to access the collection.
• new /config endpoint allows users to see what configurations are active on the visionbox docker container. This includes being able to see the changes made with the 'on-the-fly' facial-recognition model-switching.
• system now accesses a default configuration if there is no tfsdk_options.json file in the local collections volume.
• fixed an issue where tfsdk_options.json was not always being read properly when in the local collections volume.
• /face-detect endpoint updated to use the new landmark function, significantly increasing precision and accuracy.
• /liveness endpoint now takes a video rather than an image. This allows visionbox to test to see if there is an acutal blink rather than just checking if someone's eyes are opened or closed with a single image.
• Add logout functionality. Once logged out, that user's temporary token will be rejected, and a new one will be generated when they log in again.
• Removed tfsdk.FACEDETECTIONMODE from tfsdk_options.json
• /active-spoof endpoint - renamed input params and refactor response to be user friendly
• Converted /face-detect to return faceprint instead of base64 chip.
• Changed default face recognition model from LITE to TFV5. This may slow down things a bit, but it significantly increases facial-recognition accuracy. https://reference.trueface.ai/cpp/dev/latest/py/faq.html#what-are-the-differences-between-the-face-recognition-models⁠
• fixed an issue where it is possible to register 1 user with 'phantom credentials'
• fixed an issue where if num_candidates is set to zero, it causes problems
• Added a step to check if the namespace exists before loading a collection - this allows for much more accuracy in tracking down a problem as to whether the namespace is wrong or the collection name is wrong should an error occur.

0.8.1:

• Added batch identification in the /identify and /identify/zipfile endpoints.
• bugfix: /config endpoint now supports mapped volumes.
• bugfix: bugfix: handle num_candidates=0.
• Added a new spoof detection method. The /active-spoof endpoint takes two images, one close-up of the face and one further away.

0.8.0:

• Migrated MySQL user db to SQLite, in case of issues you can continue to use the last version of the visionbox (0.7.6.1) with your mysql user database. This only affects user authentication (the /register and /login endpoints), collections were already stored in SQLite so no changes there.
• /identify can now identify all faces present in the image, not just the largest one, using the identify_all_faces=True flag. Defaults to False.
• added an /identify/zipfile endpoint which identifies all images in a zipfile in the request instead of just one
• Added support for all backend SDK configuration options as described here: https://reference.trueface.ai/cpp/dev/latest/py/general.html?highlight=options#tfsdk.ConfigurationOptions⁠
  • This allows for example the selection of specific Face Recognition models like TFv5 which is our latest NIST submission model.
• Added POSTGRESQL dbms support for centralized collection databases.

0.7.6.1:

• start with one worker process by default unless otherwise specified, this helps avoid GPU memory starvation

0.7.6:

• making OAUTH2 authentication an opt-in feature. The visionbox will now only restrict access for unauthorized users once a first user has been defined.
• bugfix: deleting existing collections
• bugfix: clearer error responses on /spdetect and /match

0.7.5.1:

• Fixed bug in /match not comparing 2 urls correctly

0.7.5:

• Fix refresh token expiring before access token
• Added /namespaces [GET, DELETE] endpoint
• Added /collections [GET, DELETE] endpoint
• Relax root endpoint to allow non versioned request
• Fix /user [GET]; get all users not working
• /identify now returns n top candidates with the default being num_candidates=3

0.7.4:

• bugfix release, restoring databases on startup
• allow setting of a timeout value on /enroll

0.7.3:

• bugfix release

0.7.2:

• bugfix release
• adding support for CPUs without AVX2
• adding support to automatically make use of a GPU if present

0.7.1:

• adding feature to pass a face feature to /identify

0.7:

• internal switch to C++ SDK
• improved face detection
• improved face recognition model
• improved speed through vector compression
• collection backend changed from npz files to SQLite
• user management and authorization

0.6.6.7:

• bugfix in loading pickled collections

0.6.6.6:

• adding request logging

0.6.6.5:

• new and improved face recognition model

0.6.6.4:

• expose threshold parameter in /identify endpoint

0.6.6.3:

• improved spoof detection model

0.6.6.2:

• allows connecting to a memsql backend database

0.6.6.1:

• api can now auto recover from cuda out of memory error

0.6.5.2:

• allow source_feature to be used in /match endpoint