Trueface overhauls face biometrics SDK
A blog post by senior computer vision software developer Cyrus Behroozi states that the new release includes a TFV6 facial recognition model that records a 24 percent improvement in accuracy for masked probe images compared to the TFV5, according to the NIST FRVT. Behroozi writes that the model will work for situations where the probe image contains a masked face for 1:N searches or for 1:1 comparisons of masked faces.
The GPU SDK will use TensorRT for inference to optimize performance for Nvidia GPUs, which reduces latency while increasing throughput with more control over design such as allowing the user to choose their precision level, optimal batch size, and maximum GPU memory usage.
The SDK will no longer store images internally and will instead pass the pre-processed image to each inference function in a change known as ‘remove state’. Behroozi says this is the “most significant change made to our API to date,” and was modified due to serious design limitations that would require users to be wary of thread safety. It also allows for more versatile design patterns.
The 1.0 version will also sport an improved passive spoof model to catch presentation attacks and with dynamic library support for Linux.