⚡Fast_Seg⚡

This repo try to implement state-of-art fast semantic segmentation model s on road scene dataset(CityScape, Mapillary, Camvid).

What is purpose of this repo?

This repo aims to do experiments and verify the idea of fast semantic segmentation and this repo also provide some fast models.

Our ICnet implementation achieves 74.5% mIoU which is 5% point higher than orginal paper. !!!!! Here: model

GALD-net provides some state-of-art accurate methods implementation.

ICNet:ICnet for real-time semantic segmentation on high-resolution images. ECCV-2018, paper
DF-Net: Partial Order Pruning: for Best Speed/Accuracy Trade-off in Neural Architecture Search.CVPR-2019, paper
Bi-Seg: Bilateral segmentation network for real-time semantic segmentation.ECCV-2018, paper
DFA-Net: Deep feature aggregation for real-time semantic segmentation.CVPR-2019,paper
ESP-Net: Efficient Spatial Pyramid of Dilated Convolutions for Semantic Segmentation. ECCV-2018,paper
SwiftNet: In defense of pre-trained imagenet architectures for real-time semantic segmentation of road-driving images. CVPR2019, paper
Real-Time Semantic Segmentation via Multiply Spatial Fusion Network.(face++) arxiv,paper
Fast-SCNN: Fast Semantic Segmentation Network.BMVC-2019 paper

use train_distribute.py for training For example, use scripts in exp floder for training and evaluation.
use prediction_test_different_size.py for prediction with different size input.

You can download [cityscapes] dataset (https://www.cityscapes-dataset.com/) from here. Note: please download leftImg8bit_trainvaltest.zip(11GB) and gtFine_trainvaltest(241MB).
You can download camvid dataset from here.
You can download pretrained XceptionA with RGB input and ResNet18 with bgr input and ResNet50 with bgr input [link]:(https://pan.baidu.com/s/1mM_Lc44iX9CT1nPq6tjOAA) password:bnfv. or ['link']: resnet50-deep.pth, icnet_final.pth, resnet18-deep-caffe.pth, xceptiona_imagenet.pth

use syn-bn(apex).
use batch-size >=8.
use deep supervised loss for easier optimation.
use large crop size during training.
longer training time for small models(60,000 interaction or more).
use Mapillary data for pretraining for boosting performance.
Deeply based resnet runs slowly than torch pretrained resnet but with higher accuracy.
The small network doesn't need ImageNet pretraining if training longer time on Cityscape.(Fast-SCNN paper)