TinaFace(2021)

• Here is the Paper of the model
• I implement this model with TensorFlow version '2.6.0'

Setup

• Download my data folder including image folders for train, validate and test also label in JSON format; unzip and put it at same level with model folder. Dataset

Dataset

• My dataset is a small part of WIDER FACE dataset
• My dataset consists 6 folders: 3 folder of image 3 folder of label corresponding each image
• I use Labelme application to annotate image by myself and you can get more detail here

How this project works

• Input Pipeline: Transform raw image and bounding box coordinates in image to tensorflow dataset

  • After apply augmentation, I tranform image to 640x640x3 tensor
  • Because I don't know exactly the number of face in image so I create a "padded_bboxes" tensor with shape(100x4) then add bound box corrdinate tensor later

• Model:

  • ResNet50 --> FCN --> Inception --> Head

    • From C2 -> C3 -> C4 -> C5, the size of block reduce respectively 160x160 -> 80x80 -> 40x40 -> 20x20
    • ResNet50 block use to get the overview of image but ignore detail
    • ResNet50 combine FCN in order to get both detail and overview of the image
    • Inception block help model to capture objects as well as context at multiple scale

  • I simplify model when compare with model architecture in paper by reducing 7 levels to 5 levels in Resnet50. The deeper Resnet50 is the better model can detect close face. This is the first version so I will update model later.

• Target Assigner: Transform raw image with bounding box coordinates to heads

  • Architecture of head for each level:
    • Classification head
    • Regression head
    • IoU Aware head

Model Architecture