MatterPort3D 数据集 | 简介 | 多途径下载

Matterport3D，简称MP3D，是室内场景的一个大规模 RGB-D 数据集。可用于机器人具身导航算法开发。

包含 90 个建筑规模场景的 194,400 张RGB-D图像，以及10,800个全景视图。

数据集提供了表面重建、相机姿态以及二维和三维语义分割的注释。

支持各种监督和自监督计算机视觉任务，包括关键点匹配、视图重叠预测、基于颜色的法线预测、语义分割和场景分类。

项目地址：https://niessner.github.io/Matterport/

代码地址：https://github.com/niessner/Matterport

论文地址：Matterport3D: Learning from RGB-D Data in Indoor Environments

1、数据集下载（官方）

该数据集包含多种类型的标注：彩色和深度图像、相机姿态、带纹理的 3D 网格、建筑平面图和区域标注、对象实例语义标注。

数据集的格式参考：https://github.com/niessner/Matterport/blob/master/data_organization.md

需要填写并签署使用条款协议表格，然后将其发送至matterport3d@googlegroups.com以请求访问数据集。

2、Habitat-Lab 提供示例下载

Scenes datasets 如下表格所示：

点击“Matterport3D”后，能看到下载命令：

python -m habitat_sim.utils.datasets_download --uids mp3d_example_scene --data-path data/

但是这只是一个示例场景，用于在 Habitat-sim 中执行单元测试，无法下载完整的“Matterport3D”数据。

Task datasets 如下表格所示：

3、批量下载

首先创建一个Conda环境，名字为mp3d，python版本为3.9

进入mp3d环境

conda create -n mp3d python=3.9
conda activate mp3d

然后安装requests依赖库

pip install requests

编写批量下载的代码：

#!/usr/bin/env python
# -*- coding: utf-8 -*-

import argparse
import collections
import os
import tempfile
import urllib
from urllib import request

import requests
from time import sleep
from requests.exceptions import ConnectionError, ChunkedEncodingError, RequestException
import sys


BASE_URL = 'http://kaldir.vc.in.tum.de/matterport/'
RELEASE = 'v1/scans'
RELEASE_TASKS = 'v1/tasks/'
RELEASE_SIZE = '1.3TB'
TOS_URL = BASE_URL + 'MP_TOS.pdf'
FILETYPES = [
    'cameras',
    'matterport_camera_intrinsics',
    'matterport_camera_poses',
    'matterport_color_images',
    'matterport_depth_images',
    'matterport_hdr_images',
    'matterport_mesh',
    'matterport_skybox_images',
    'undistorted_camera_parameters',
    'undistorted_color_images',
    'undistorted_depth_images',
    'undistorted_normal_images',
    'house_segmentations',
    'region_segmentations',
    'image_overlap_data',
    'poisson_meshes',
    'sens'
]
TASK_FILES = {
    'keypoint_matching_data': ['keypoint_matching/data.zip'],
    'keypoint_matching_models': ['keypoint_matching/models.zip'],
    'surface_normal_data': ['surface_normal/data_list.zip'],
    'surface_normal_models': ['surface_normal/models.zip'],
    'region_classification_data': ['region_classification/data.zip'],
    'region_classification_models': ['region_classification/models.zip'],
    'semantic_voxel_label_data': ['semantic_voxel_label/data.zip'],
    'semantic_voxel_label_models': ['semantic_voxel_label/models.zip'],
    'minos': ['mp3d_minos.zip'],
    'gibson': ['mp3d_for_gibson.tar.gz'],
    'habitat': ['mp3d_habitat.zip'],
    'pixelsynth': ['mp3d_pixelsynth.zip'],
    'igibson': ['mp3d_for_igibson.zip'],
    'mp360': ['mp3d_360/data_00.zip', 'mp3d_360/data_01.zip', 'mp3d_360/data_02.zip', 'mp3d_360/data_03.zip', 'mp3d_360/data_04.zip', 'mp3d_360/data_05.zip', 'mp3d_360/data_06.zip']
}


def get_release_scans(release_file):
    # scan_lines = urllib.urlopen(release_file)
    scan_lines = request.urlopen(release_file)
    scans = []
    for scan_line in scan_lines:
        scan_line = str(scan_line, 'utf-8')
        scan_id = scan_line.rstrip('\n')
        scans.append(scan_id)
    return scans


def download_release(release_scans, out_dir, file_types):
    print('Downloading MP release to ' + out_dir + '...')
    for scan_id in release_scans:
        scan_out_dir = os.path.join(out_dir, scan_id)
        download_scan(scan_id, scan_out_dir, file_types)
    print('Downloaded MP release.')


def download_file(url, out_file, max_retries=5, chunk_size=1024*1024):
    """
    下载单个文件，支持断点续传和自动重试，并打印实时下载进度。
    兼容 Python2/3。
    """
    out_dir = os.path.dirname(out_file)
    if not os.path.isdir(out_dir):
        try:
            os.makedirs(out_dir)
        except OSError:
            pass

    # 获取文件总大小
    head = requests.head(url, allow_redirects=True)
    if head.status_code != 200:
        raise IOError("无法获取文件大小: {0}，状态码 {1}".format(url, head.status_code))
    total_size = int(head.headers.get('Content-Length', 0))

    # 计算已下载起点
    resume = 0
    if os.path.exists(out_file):
        resume = os.path.getsize(out_file)

    if resume >= total_size:
        print("跳过已存在文件 %s" % out_file)
        return

    print("开始下载 %s (%0.2f MB)" % (out_file, total_size / 1024.0**2))

    retries = 0
    last_print = 0
    while resume < total_size and retries <= max_retries:
        headers = {"Range": "bytes=%d-%d" % (resume, total_size - 1)}
        try:
            r = requests.get(url, headers=headers, stream=True, timeout=30)
            r.raise_for_status()

            fh, tmp_path = tempfile.mkstemp(dir=out_dir)
            with os.fdopen(fh, "wb") as tmpf:
                for chunk in r.iter_content(chunk_size=chunk_size):
                    if not chunk:
                        continue
                    tmpf.write(chunk)
                    resume += len(chunk)

                    # 每下载 5% 或最后一块时打印一次进度
                    percent = int(resume * 100 / total_size)
                    if percent - last_print >= 5 or resume == total_size:
                        sys.stdout.write("\r下载进度: %3d%% (%0.2f/%0.2f MB)" % (
                            percent,
                            resume  / 1024.0**2,
                            total_size / 1024.0**2
                        ))
                        sys.stdout.flush()
                        last_print = percent

            # 将本次下载数据追加到目标文件
            with open(tmp_path, "rb") as tmpf, open(out_file, "ab") as outf:
                outf.write(tmpf.read())
            os.remove(tmp_path)
            r.close()
            break

        # except (requests.ConnectionError, requests.ChunkedEncodingError, IOError) as e:
        except (ConnectionError, ChunkedEncodingError, IOError) as e:
            retries += 1
            wait = 2 ** retries
            print("\n[重试 %d/%d] 已下载 %0.2f MB，等待 %d 秒后重试..." % (
                retries, max_retries,
                resume / 1024.0**2,
                wait
            ))
            sleep(wait)

    if resume < total_size:
        raise IOError("下载失败：只获取到 %d / %d 字节" % (resume, total_size))

    # 下载完成后换行
    sys.stdout.write("\n下载完成：%s\n" % out_file)


def download_scan(scan_id, out_dir, file_types):
    print('Downloading MP scan ' + scan_id + ' ...')
    if not os.path.isdir(out_dir):
        os.makedirs(out_dir)
    for ft in file_types:
        url = BASE_URL + RELEASE + '/' + scan_id + '/' + ft + '.zip'
        out_file = out_dir + '/' + ft + '.zip'
        download_file(url, out_file)
    print('Downloaded scan ' + scan_id)


def download_task_data(task_data, out_dir):
    print('Downloading MP task data for ' + str(task_data) + ' ...')
    for task_data_id in task_data:
        if task_data_id in TASK_FILES:
            file = TASK_FILES[task_data_id]
            for filepart in file:
                url = BASE_URL + RELEASE_TASKS + '/' + filepart
                localpath = os.path.join(out_dir, filepart)
                localdir = os.path.dirname(localpath)
                if not os.path.isdir(localdir):
                    os.makedirs(localdir)
                    download_file(url, localpath)
                    print('Downloaded task data ' + task_data_id)


def main():
    parser = argparse.ArgumentParser(description=
        '''
        Downloads MP public data release.
        Example invocation:
          python download_mp.py -o base_dir --id ALL --type object_segmentations --task_data semantic_voxel_label_data semantic_voxel_label_models
        The -o argument is required and specifies the base_dir local directory.
        After download base_dir/v1/scans is populated with scan data, and base_dir/v1/tasks is populated with task data.
        Unzip scan files from base_dir/v1/scans and task files from base_dir/v1/tasks/task_name.
        The --type argument is optional (all data types are downloaded if unspecified).
        The --id ALL argument will download all house data. Use --id house_id to download specific house data.
        The --task_data argument is optional and will download task data and model files.
        ''',
        formatter_class=argparse.RawTextHelpFormatter)
    parser.add_argument('-o', '--out_dir', required=True, help='directory in which to download')
    parser.add_argument('--task_data', default=[], nargs='+', help='task data files to download. Any of: ' + ','.join(TASK_FILES.keys()))
    parser.add_argument('--id', default='ALL', help='specific scan id to download or ALL to download entire dataset')
    parser.add_argument('--type', nargs='+', help='specific file types to download. Any of: ' + ','.join(FILETYPES))
    args = parser.parse_args()


    release_file = BASE_URL + RELEASE + '.txt'
    release_scans = get_release_scans(release_file)
    file_types = FILETYPES

    # download task data
    if args.task_data:
        if set(args.task_data) & set(TASK_FILES.keys()):  # download task data
            out_dir = os.path.join(args.out_dir, RELEASE_TASKS)
            download_task_data(args.task_data, out_dir)
        else:
            print('ERROR: Unrecognized task data id: ' + args.task_data)
        print('Done downloading task_data for ' + str(args.task_data))
        # key = raw_input('Press any key to continue on to main dataset download, or CTRL-C to exit.')
        key = input('Press any key to continue on to main dataset download, or CTRL-C to exit.')

    # download specific file types?
    if args.type:
        if not set(args.type) & set(FILETYPES):
            # print('ERROR: Invalid file type: ' + file_type)
            print('ERROR: Invalid file type: ' + file_types)
            return
        file_types = args.type

    if args.id and args.id != 'ALL':  # download single scan
        scan_id = args.id
        if scan_id not in release_scans:
            print('ERROR: Invalid scan id: ' + scan_id)
        else:
            out_dir = os.path.join(args.out_dir, RELEASE, scan_id)
            download_scan(scan_id, out_dir, file_types)
    elif 'minos' not in args.task_data and args.id == 'ALL' or args.id == 'all':  # download entire release
        if len(file_types) == len(FILETYPES):
            print('WARNING: You are downloading the entire MP release which requires ' + RELEASE_SIZE + ' of space.')
        else:
            print('WARNING: You are downloading all MP scans of type ' + file_types[0])
        print('Note that existing scan directories will be skipped. Delete partially downloaded directories to re-download.')
        print('***')
        print('Press any key to continue, or CTRL-C to exit.')
        # key = raw_input('')
        key = input('')
        out_dir = os.path.join(args.out_dir, RELEASE)
        download_release(release_scans, out_dir, file_types)

if __name__ == "__main__": main()

创建一个文件夹，用于存放数据集，比如data

下载所以MP3D的数据，1.3T左右，执行命令：

python download_mp.py --task habitat -o ./data

打印信息：

Downloading MP task data for ['habitat'] ...
Done downloading task_data for ['habitat']
Press any key to continue on to main dataset download, or CTRL-C to exit.
WARNING: You are downloading the entire MP release which requires 1.3TB of space.
Note that existing scan directories will be skipped. Delete partially downloaded directories to re-download.
***
Press any key to continue, or CTRL-C to exit.

Downloading MP release to ./data/v1/scans...

Downloading MP scan 1pXnuDYAj8r ...
下载完成：./data/v1/scans/1pXnuDYAj8r/cameras.zip
开始下载 ./data/v1/scans/1pXnuDYAj8r/matterport_camera_intrinsics.zip (0.10 MB)
下载进度: 100% (0.10/0.10 MB)
下载完成：./data/v1/scans/1pXnuDYAj8r/matterport_camera_intrinsics.zip
开始下载 ./data/v1/scans/1pXnuDYAj8r/matterport_camera_poses.zip (0.61 MB)
下载进度: 100% (0.61/0.61 MB)
下载完成：./data/v1/scans/1pXnuDYAj8r/matterport_camera_poses.zip
开始下载 ./data/v1/scans/1pXnuDYAj8r/matterport_color_images.zip (537.00 MB)
下载进度: 100% (537.00/537.00 MB)
............
下载完成：./data/v1/scans/1pXnuDYAj8r/undistorted_color_images.zip
开始下载 ./data/v1/scans/1pXnuDYAj8r/undistorted_depth_images.zip (1282.63 MB)
下载进度: 100% (1282.63/1282.63 MB)
下载完成：./data/v1/scans/1pXnuDYAj8r/undistorted_depth_images.zip
开始下载 ./data/v1/scans/1pXnuDYAj8r/undistorted_normal_images.zip (2808.76 MB)
下载进度:  90% (2528.00/2808.76 MB)
 

下载过程中，场景数据的文件：

17DRP5sb8fy任务包含的内容：

...............

任务数据文件：

指定某个场景ID，进行下载，执行命令：

python download_mp.py --task habitat -o ./data --id 17DRP5sb8fy

打印信息：

Downloading MP scan 17DRP5sb8fy ...
开始下载 ./data/v1/scans/17DRP5sb8fy/cameras.zip (0.00 MB)
下载进度: 100% (0.00/0.00 MB)
下载完成：./data/v1/scans/17DRP5sb8fy/cameras.zip
开始下载 ./data/v1/scans/17DRP5sb8fy/matterport_camera_intrinsics.zip (0.05 MB)
下载进度: 100% (0.05/0.05 MB)
........
下载完成：./data/v1/scans/17DRP5sb8fy/image_overlap_data.zip
开始下载 ./data/v1/scans/17DRP5sb8fy/poisson_meshes.zip (136.50 MB)
下载进度: 100% (136.50/136.50 MB)
下载完成：./data/v1/scans/17DRP5sb8fy/poisson_meshes.zip
开始下载 ./data/v1/scans/17DRP5sb8fy/sens.zip (1402.69 MB)
下载进度: 100% (1402.69/1402.69 MB)
下载完成：./data/v1/scans/17DRP5sb8fy/sens.zip
Downloaded scan 17DRP5sb8fy

4、便捷下载（简要文件）

 下载地址：https://cloud.tsinghua.edu.cn/f/03e0ca1430a344efa72b/?dl=1

每个文件夹中，包含一个.glb文件：

分享完成～