Skip to content

uber-research/UPSNet

Repository files navigation

UPSNet: A Unified Panoptic Segmentation Network

Introduction

UPSNet is initially described in a CVPR 2019 oral paper.

Disclaimer

This repository is tested under Python 3.6, PyTorch 0.4.1. And model training is done with 16 GPUs by using horovod. It should also work under Python 2.7 / PyTorch 1.0 and with 4 GPUs.

License

© Uber, 2018-2019. Licensed under the Uber Non-Commercial License.

Citing UPSNet

If you find UPSNet is useful in your research, please consider citing:

@inproceedings{xiong19upsnet,
    Author = {Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, Raquel Urtasun},
    Title = {UPSNet: A Unified Panoptic Segmentation Network},
    Conference = {CVPR},
    Year = {2019}
}

Main Results

COCO 2017 (trained on train-2017 set)

test split PQ SQ RQ PQTh PQSt
UPSNet-50 val 42.5 78.0 52.4 48.5 33.4
UPSNet-101-DCN test-dev 46.6 80.5 56.9 53.2 36.7

Cityscapes

PQ SQ RQ PQTh PQSt
UPSNet-50 59.3 79.7 73.0 54.6 62.7
UPSNet-101-COCO (ms test) 61.8 81.3 74.8 57.6 64.8

Requirements: Software

We recommend using Anaconda3 as it already includes many common packages.

Requirements: Hardware

We recommend using 4~16 GPUs with at least 11 GB memory to train our model.

Installation

Clone this repo to $UPSNet_ROOT

Run init.sh to build essential C++/CUDA modules and download pretrained model.

For Cityscapes:

Assuming you already downloaded Cityscapes dataset at $CITYSCAPES_ROOT and TrainIds label images are generated, please create a soft link by ln -s $CITYSCAPES_ROOT data/cityscapes under UPSNet_ROOT, and run init_cityscapes.sh to prepare Cityscapes dataset for UPSNet.

For COCO:

Assuming you already downloaded COCO dataset at $COCO_ROOT and have annotations and images folders under it, please create a soft link by ln -s $COCO_ROOT data/coco under UPSNet_ROOT, and run init_coco.sh to prepare COCO dataset for UPSNet.

Training:

python upsnet/upsnet_end2end_train.py --cfg upsnet/experiments/$EXP.yaml

Test:

python upsnet/upsnet_end2end_test.py --cfg upsnet/experiments/$EXP.yaml

We provide serveral config files (16/4 GPUs for Cityscapes/COCO dataset) under upsnet/experiments folder.

Model Weights

The model weights that can reproduce numbers in our paper are available now. Please follow these steps to use them:

Run download_weights.sh to get trained model weights for Cityscapes and COCO.

For Cityscapes:

python upsnet/upsnet_end2end_test.py --cfg upsnet/experiments/upsnet_resnet50_cityscapes_16gpu.yaml --weight_path ./model/upsnet_resnet_50_cityscapes_12000.pth
python upsnet/upsnet_end2end_test.py --cfg upsnet/experiments/upsnet_resnet101_cityscapes_w_coco_16gpu.yaml --weight_path ./model/upsnet_resnet_101_cityscapes_w_coco_3000.pth

For COCO:

python upsnet/upsnet_end2end_test.py --cfg upsnet/experiments/upsnet_resnet50_coco_16gpu.yaml --weight_path model/upsnet_resnet_50_coco_90000.pth
python upsnet/upsnet_end2end_test.py --cfg upsnet/experiments/upsnet_resnet101_dcn_coco_3x_16gpu.yaml --weight_path model/upsnet_resnet_101_dcn_coco_270000.pth