This repo is about face segmentation based on LinkNet34, our work is inspired by this PyTorch implementaion.
Spec | |
---|---|
Operating System | Ubuntu 18.04 |
GPU | Nvidia A100-SXM4-40GB |
CUDA Version | 11.0 |
Driver Version | 460.73.01 |
Oneflow Version | branch: master, commit_id: 90d3277a098f483d0a0e68621b7c8fb2497a6fc2 |
pip3 install -r requirements.txt
wget https://oneflow-public.oss-cn-beijing.aliyuncs.com/model_zoo/cv/seg/linknet_oneflow_model.zip
bash infer.sh
We combine several dataset and preprocess them, for training and evaluating our
face segmentation model. We provide the final compressed dataset, please download dataset from
here , unzip it and put it in ./
Please download pretrained ResNet backbone from pretrained_resnet_oneflow_model
, unzip it and put it in ./
Hyperparameters can be customized, the provided model is trained by run
bash train.sh
We compare our model with baseline implemented using PyTorch, our model on OneFlow performs even little bit better on train dataset than baseline. The result is shown in the following,
Train IoU (%) | Test IoU (%) | |
---|---|---|
PyTorch | 92.815 | 92.689 |
OneFlow | 95.006 | 94.425 |
The following displays some segmented results on grabbed videos.