Channel-wise Distillation for Semantic Segmentation

Introduction

This repository contains the PyTorch test implementation of: Channel-wise Distillation for Semantic Segmentation

Requirements

All the codes are tested in the following environment:

Linux (tested on Ubuntu 16.04 / CentOS 7.6)
Python 3.6.2
PyTorch 0.4.1
Single TITAN Xp GPU

Installation

Install PyTorch: conda install pytorch=0.4.1 cuda90 torchvision -c pytorch
Install other dependences: pip install opencv-python scipy
Install InPlace-ABN:

cd libs
sh build.sh
python build.py

The build.sh script assumes that the nvcc compiler is available in the current system search path. The CUDA kernels are compiled for sm_50, sm_52 and sm_61 by default. To change this (e.g. if you are using a Kepler GPU), please edit the CUDA_GENCODE variable in build.sh.

Dataset & Models

Dataset: [Cityscapes]
After distillation: PSPNet (ResNet-18) rn18-cityscape_singleAndWhole_val-75.05_test-73.86.pth [Google Drive] rn18-cityscape_singleAndWhole_val-75.90_test-74.58.pth [Google Drive]

Please create a new folder ckpt and move all downloaded models to it.

Usage

1. Inference with the evaluation dataset

python valandtest.py --data-dir path/to/CITYSCAPES/data/cityscapes  --restore-from ckpt/new_rn18-cityscape_singleAndWhole_val-75.02_test-73.86.pth --gpu 0 --type val  --figsavepath 75.02val

python valandtest.py --data-dir path/to/CITYSCAPES/data/cityscapes  --restore-from ckpt/new_rn18-cityscape_singleAndWhole_val-75.90_test-74.58.pth --gpu 0 --type val  --figsavepath 75.90val

2. Inference on the test dataset

python valandtest.py --data-dir path/to/CITYSCAPES/data/cityscapes  --restore-from ckpt/new_rn18-cityscape_singleAndWhole_val-75.02_test-73.86.pth --gpu 0 --type test --figsavepath 73.86test

python valandtest.py --data-dir path/to/CITYSCAPES/data/cityscapes  --restore-from ckpt/new_rn18-cityscape_singleAndWhole_val-75.90_test-74.58.pth --gpu 0 --type test --figsavepath 74.58test

Model	Average	roda	sidewalk	building	wall	fence	pole	trafficlight	trafficsign	vegetation	terrain	sky	person	rider	car	truck	bus	train	motorcycle	bicycle
IoU	73.86	97.84	81.01	91.55	48.63	53.18	61.14	70.21	74.20	92.93	70.91	94.84	83.11	62.39	94.74	54.12	66.80	70.91	61.60	73.27
IoU	74.58	97.78	80.56	91.45	52.78	52.91	59.90	70.50	73.13	92.54	70.70	94.57	82.25	63.51	94.76	59.31	73.68	73.00	61.54	72.12

Citation

Please consider citing this work if it helps your research:


@inproceedings{shu2020cwd,
  title={Channel-wise Distillation for Semantic Segmentation},
  author={Shu, Changyong and Liu, Yifan and Gao, Jianfei and Xu, Lin and Shen, Chunhua},
}

irfanICMLL/CWD