Please enable JavaScript.

Coggle requires JavaScript to display documents.

Face Recognition (Mobile (ShuffleNet (Object Detection Task from MSCOCO…

- - - - Using FaceNet (distilation) #
    - - Tensorflow
        
        Quantized versions
        
        Benchmark on github page
        
        Fast forward on CPU is 19ms
      - pytorch
        
        It have a benchmark
      - Official on tensorflow
  - - - 1.0 MobileNet-224
        
        70.6% accuracy
        
        4.2M parameters
        
        569M mult-adds
      - VGG16
        
        71.5% accuracy
        
        138M parameters
        
        15300M mult-adds
      - SqueezeNet
        
        57.5% accuracy
        
        1.25M Parameters
        
        1700M mult-adds
      - 0.5 MobileNet-160
        
        60.2% accuracy
        
        1.32M Parameters
        
        76M mult-adds
      - AlexNet
        
        57.2% accuracy
        
        60M
        
        720M mult-adds
      - Source: MobileNet Paper
      - ShuffeNet 2x (g=3)
        
        73.7% accuracy
        
        524M mult-adds
      - source: ShuffleNet Paper
      - ShuffleNet 1 (g=8)
        
        67.6% accuracy
        
        140M mult-adds
  - - - Faster-RCNN based
      - Better results than MobileNet based Faster-RCNN
    - - Inference time table
        
        224x224
        
        Shuffle: 15.2-108.0ms
        
        1.0 MobileNet-224: 110.0ms
        
        720x1280
        
        Shuffle: 260.1-1857.6ms
        
        1.0 MobileNet-224: 1879
        
        Snapdragon 820 processor, single thread
    - - caffe
      - tensorflow
- - - - CASIA-WebFace
        
        LFW accuracy: 0.987
      - Subset of MS-Celeb-1M
        
        Best trainings
        
        LFW accuracy: 0.992
    - - MTCNN
        
        Detection
        
        Alignment
    - - ResNet V1
      - ResNet V2
      - SqueezeNet