Experiment plan

[eddogola]

First run of experiments:

Big picture question - How much efficiency can we squeeze out of UNet and adjacent(MaskR-CNN+faster variants, SSD, MobileNet, FCN, FPN) models without losing on accuracy?

• Experiment 1 - Learn/get a model on browser using ONNX [yolov8 on onnx]
• Experiment 2 - Train, prune, quantize model from scratch
  • U-Net [U-Net from scratch]
  • Squeeze U-Net [Squeeze U-Net from scratch]
  • SSD
• Experiment 3 - Fine-tune pre-trained models or ones with a pre-trained backbone
  • U-Net (with pre-trained resnet50) [U-Net with resnet50 backbone]
  • MaskR-CNN**
    • [RCNN on onnx]
  • SSD
  • MobileNet
  • Yolov8 [yolov8 on onnx]
• Experiment 5 - Porting an implementation to C/C++ and running using webassembly
• Experiment 4 - Compare efficiency, accuracy among models, and model limits(image size, throughput)