T-Top - DNN improvements

[mamaheux]

• Use AdamW
• Use knowledge distillation
• Object detection
  • Use pretrained yolov7 and yolov7tiny
  • Use yolov7 and yolov7tiny for object descriptor
• Person pose
  • Use efficient net instead of resnet18
  • Optimize the gaussian parameters
  • Increase the resolution of the output
  • BCE divergence for knowledge distillation
  • Training
• Face recognition
  • Improve the face alignment, maybe just a crop
  • Use efficient net instead of OpenFace
  • Use ArcFace instead AM-Softmax (https://openaccess.thecvf.com/content_CVPR_2019/papers/Deng_ArcFace_Additive_Angular_Margin_Loss_for_Deep_Face_Recognition_CVPR_2019_paper.pdf)
  • https://odr.chalmers.se/server/api/core/bitstreams/0f358a79-f2e1-48f3-b591-a4563489cd8f/content section 3.3 for knowledge distillation
  • Test also on a modern dataset like
  • Training
• Audio classification
  • Use AudioSet
  • Use a smaller frame size
  • Use https://arxiv.org/pdf/2102.01243v3.pdf
  • Training
• Speaker identification
  • Reduce the frame duration to 1 sec
  • Use the full size network
  • Use the pretrained network from speechbrain
  • Training