search

anilsathyan7 / Portrait-Segmentation

Real-time portrait segmentation for mobile devices
MIT License
638 stars 133 forks source link

Increase segmentation accuracy #28

Closed ghost closed 3 years ago

ghost commented 3 years ago

Hi, thank you very much for your previous help. How can I significantly increase segmentation accuracy of the slim512 network, speed is not important in principle, I need a very high segmentation accuracy and the ability to convert the model to TFLite.

anilsathyan7 commented 3 years ago

There's no silver bullet; you need to experiment on various parameter settings and compare their results. Anyway, here are some suggestions:-

  1. May be use better and bigger datasets/ input image size to suit your use case.
  2. Use advanced training techniques like boundary loss, temporal consistency etc.
  3. Make the model deeper/ bigger by adding layers, increasing number of filters etc.

Also try out heavier models like deeplab, bisnet, boundary-aware network etc. if speed is not an issue...

ghost commented 3 years ago

Thanks, I probably won't be able to do it, I'll try to figure it out.

ghost commented 3 years ago

How can I for example use resnet18 or vgg16 to train the model?

anilsathyan7 commented 3 years ago

I would suggest you to look at the working of basic encoder decoder networks and how to use popular networks like mobilenet or resnet as backbone for segmentaion tasks.

ghost commented 3 years ago

Thanks, i use DeeplabV3Plus now