This project removes your background by optimizing segmentation models for CPU inference. It applies deep learning optimizations & engineering optimizations to improve CPU inference frame rate from 3 FPS to 15 FPS.
Original Unet with ResNet Unet with MobileNet
UNet-EfficientNet EB0 UNet-EfficientNet EB1 UNet-EfficientNet EB2
| Optimization | CPU Inference time | FPS | Model Size | IOU |
|---|---|---|---|---|
| ResNet Model (256*256) | ~265 ms | 3-4 | 98.2 MB | 0.95592 |
| Architecture Optimization (EfficientNet EB0) | ~229 ms | 5-6 | 41.2 MB | 0.96079 |
| Deep Learning Optimization (Quantization + Graph Optimization + EfficientNet EB0) | ~219 ms | 5-6 | 10.6 MB | 0.92648 |
| Engineering Optimization (Quantization + Graph Optimization + EfficientNet EB0 + Image resolution reduction - 128*128) | ~56 ms | 15-17 | 10.6 MB | 0.92648 |
The optimized and compressed models can be found under following directory
/data/assets/saved_models
Add docker image
Further details regarding the motivation, methods and results of different optimization techniques can be found in my presentation here.
Copyright (c) 2019 Chinmay Naik