This repository provides a real-time demo of ESPNetv2 on iPhone (tested only on iPhone7). Below are some illustrations.

Real-time semantic segmentation using ESPNetv2 on iPhone7
Seg demo on iPhone7 Seg demo on iPhone7

Model details

The COREML ESPNetv2 model takes an RGB image of size 256x256 as an input and produces an output of size 256x256 in real-tim. The model learns about 0.79 million parameters and performs roughly 337 million FLOPs to generate the segmentation mask. The model is trained using PyTorch on the PASCAL VOC 2012 dataset and achieves a segmentation score of 63.36, which is measured in terms of mean interesection over union (mIOU).

Several pre-trained models are provided in our EdgeNets repository.


If you are familiar with iOS application development and wants to improve the design or contribute in some way, please do so by creating a pull request. We welcome contributions.