README.md

SAM - Face recognition system (DEPLOY BRANCH)

Face recognition for identification case study written in Python using PyTorch, OpenCV, dlib

Branch details

This branch is ougt to hold all the utilities and necessary stuff to run the face recognition on the Nvidia Jetson TX1 board, usually referred to as the edge.

Requirements:

• WebCam
• A host (for us it is TX1) capable of compiling the following:
  • OpenCV - Video capture
  • Dlib - Face alignment and detection
  • PyTorch - Recognition neural network
• A DEPLOY_DATABASE.tar that holds facial embeddings and corresponding data for authorization

Latest benchmarks

Embedding network definition:

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.features = models.squeezenet1_1().features
        self.embedding = nn.Sequential(
            nn.Linear(2048, 512),
            nn.Dropout(),
            nn.ReLU(),
            nn.Linear(512, 128)
        )

    def forward(self, x):
        x = self.features(x)
        x = nn.functional.adaptive_max_pool2d(x, 2)
        batch_size = x.size(0)
        x = x.view(batch_size, -1)
        return self.embedding(x)

Training results

	Training set	Test set
# Individuals	467	41248
# Images	88	10000
Accuracy (top-1)	99.6%	99.3%

Time complexity

TASK	P100	K80	Jetson TX1
640x480x3 face detection (CPU)	55.1 ms ± 9.26 µs	55.1 ms ± 9.26 µs	149 ms ± 372 µs
640x480 face detection (CPU)	55.1 ms ± 9.26 µs	55.1 ms ± 9.26 µs	123 ms ± 1.17 ms
16x3x96x96 embedding net inference	3.61 ms ± 141 µs	5.08 ms ± 6.7 µs	19.1 ms ± 967 µs
1x3x96x96 embedding net inference	3.5 ms ± 112 µs	3.56 ms ± 133 µs	10.9 ms ± 358 µs
K-Nearest Neighbour from 10000x128 embeddings	386 µs ± 112 ns	648 µs ± 194 ns	2.66 ms ± 14.8 µs