/OmniCity-v1.0

The models, datasets(satellite&street view) and correlative config files of OmniCity-v1.0 project.

Primary LanguagePythonMIT LicenseMIT

OmniCity-v1.0

Introduction

This project is used in Multi-level task based on MMDetection, which includes models and correlative configs.
Note： If you want to check more about our work, please refer here.

Overview of the models and its related tasks

Tasks: Instance segmentation, Land use segmentation, Plane segmentation.

Satellite level: small view

Data Type Size Method Task Download
Satallite image 512*512 Mask R-CNN Instance segmentation Model&log
Satallite image 512*512 MS R-CNN Instance segmentation Model&log
Satallite image 512*512 CARAFE Instance segmentation Model&log
Satallite image 512*512 Cascade Instance segmentation Model&log
Satallite image 512*512 HTC Instance segmentation Model&log

Satellite level: medium view

Data Type Size Method Task Download
Satallite image 512*512 Mask R-CNN Instance segmentation Model&log

Satellite level: large view

Data Type Size Method Task Download
Satellite image 512*512 Msak R-CNN Instance segmentation Model&log

Street level: Panorama

Data Type Size Method Task Download
Panorama image 512*1024 Mask R-CNN Instance segmentation Model&log
Panorama image 512*1024 Mask R-CNN Land use segmentation Model&log
Panorama image 512*1024 MS R-CNN Land use segmentation Model&log
Panorama image 512*1024 CARAFE Land use segmentation Model&log
Panorama image 512*1024 Cascade Land use segmentation Model&log
Panorama image 512*1024 HTC Land use segmentation Model&log

Street level: Mono-view

Data Type Size Method Task Download
Mono-view image 512*512 Mask R-CNN Instance segmentation Model&log
Mono-view image 512*512 Mask R-CNN Land use segmentation Model&log
Mono-view image 512*512 Mask R-CNN Plane segmentation Model&log

Usage

Data preparation

The OmniCity dataset can be downloaded from https://opendatalab.com/OmniCity.

If you want to use your own dataset test the models above, please prepare data following MMdetection(Dataset in COCO format is preferred). And the data structure should look like below:

mmdetection
├── data
│   ├── coco
│   │   ├── annotations
│   │   ├── train2017
│   │   ├── val2017
│   │   ├── test2017

Model test

With OmniCity dataset:

# single-gpu testing
python tools/test.py \
    ${CONFIG_FILE} \
    ${CHECKPOINT_FILE} \
    [--out ${RESULT_FILE}] \
    [--eval ${EVAL_METRICS}] \
    [--show]

# CPU: disable GPUs and run single-gpu testing script
export CUDA_VISIBLE_DEVICES=-1
python tools/test.py \
    ${CONFIG_FILE} \
    ${CHECKPOINT_FILE} \
    [--out ${RESULT_FILE}] \
    [--eval ${EVAL_METRICS}] \
    [--show]

# multi-gpu testing
bash tools/dist_test.sh \
    ${CONFIG_FILE} \
    ${CHECKPOINT_FILE} \
    ${GPU_NUM} \
    [--out ${RESULT_FILE}] \
    [--eval ${EVAL_METRICS}]

With new dataset:

  • Prepare the dataset following the above rules
  • Refer to the preceding operations