TartanAir：AirSim 模擬資料集，用於同時進行當地語系化和對應 (SLAM)

發行項
09/04/2024

對於機器人來說，最基本的功能之一就是即時定位與地圖構建 (Simultaneous Localization and Mapping, SLAM)。因為影像無所不在，所以圖像式 SLAM (V-SLAM) 已成為眾多自發性系統的重要元件。幾何方法與學習方法這兩部份，已有了相當重大的進展。但為實際應用來開發耐用且可靠的 SLAM 方法，卻仍是一大挑戰。現實生活的環境充滿著困難的情況，例如光線的變化或是缺乏照明、物體並非一成不變，以及質地不明顯的情況。此資料集運用先進的電腦圖形技術，目標是全盤涵蓋模擬過程中多樣化不易達成之特性的案例。

注意

Microsoft 依「現況」提供 Azure 開放資料集。針對　貴用戶對資料集的使用方式，Microsoft 不提供任何明示或默示的擔保、保證或條件。在　貴用戶當地法律允許的範圍內，針對因使用資料集而導致的任何直接性、衍生性、特殊性、間接性、附隨性或懲罰性損害或損失，Microsoft 概不承擔任何責任。

此資料集是根據 Microsoft 接收來源資料的原始條款所提供。資料集可能包含源自 Microsoft 的資料。

人造和自然環境的影像拼貼

該資料收集自具有各式光線條件、天候與移動物體的逼真模擬環境。藉由收集模擬過程的資料，我們能夠取得多模態感應器資料與精確的實真確認標籤，包括立體 RGB 影像、深度影像、分割影像、光流與相機姿勢偵測。我們設定了大量具有各式形態與場景的環境，涵蓋了難以觀測的視角以及多樣化的移動模式，而這些內容對於使用實際資料收集平台的方式來說，實難取得。資料集的四個最重要的功能包括：1) 大量多樣化的實際資料；2) 多模式地面實況標籤；3) 動作模式的多樣性；4) 具挑戰性的場景。

此資料集提供五種資料：

立體影像：影像類型 (PNG)
深度檔案：numpy 類型 (NPY)
分割檔案：numpy 類型 (NPY)
光流檔案：numpy 類型 (NPY)
相機姿勢偵測檔案：文字類型 (TXT)

截至 2019 年止，資料集從各種不同環境中收集了共數百條軌跡 (3 TB)。

高難度視覺效果

在某些模擬情況下，資料集可模擬各種類型的高難度視覺效果。

光線條件不佳。晝夜替換。光線太暗。照明情況快速變更。
天氣效果。晴天、雨天、下雪、颳風和多霧。
季節更迭。

儲存位置

此資料集儲存於美國東部 Azure 區域。建議您在美國東部配置計算資源，以確保同質性。

授權條款

此專案透過 MIT 授權所發行。如需詳細資料，請檢閱授權檔案。

其他資訊

檢視官方 TartanAir 網站或檢視原始研究論文。

如果您對資料來源有任何疑問，請傳送電子郵件至 tartanair@hotmail.com。您也可以在相關的 GitHub 上與參與者交流。

如需詳細的引用技術資料，請參閱 AirSim paper (FSR 2017 Conference) (AirSim 論文 (FSR 2017 研討會))。請以下列方式引用：

@article{tartanair2020arxiv,
  title =   {TartanAir: A Dataset to Push the Limits of Visual SLAM},
  author =  {Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, Sebastian Scherer},
  journal = {arXiv preprint arXiv:2003.14338},
  year =    {2020}, 
  url = {https://arxiv.org/abs/2003.14338}
}

@inproceedings{airsim2017fsr,
  author = {Shital Shah and Debadeepta Dey and Chris Lovett and Ashish Kapoor},
  title = {AirSim: High-Fidelity Visual and Physical Simulation for Autonomous Vehicles},
  year = {2017},
  booktitle = {Field and Service Robotics},
  eprint = {arXiv:1705.05065},
  url = {https://arxiv.org/abs/1705.05065}
}

資料存取

使用下列程式碼範例來存取 Python 筆記本中的資料。

相依性

pip install numpy
pip install azure-storage-blob
pip install opencv-python

匯入和容器用戶端

from azure.storage.blob import ContainerClient
import numpy as np
import io
import cv2
import time
import matplotlib.pyplot as plt
%matplotlib inline

# Dataset website: http://theairlab.org/tartanair-dataset/
account_url = 'https://tartanair.blob.core.windows.net/'
container_name = 'tartanair-release1'

container_client = ContainerClient(account_url=account_url, 
                                 container_name=container_name,
                                 credential=None)

環境和軌跡

def get_environment_list():
    '''
    List all the environments shown in the root directory
    '''
    env_gen = container_client.walk_blobs()
    envlist = []
    for env in env_gen:
        envlist.append(env.name)
    return envlist

def get_trajectory_list(envname, easy_hard = 'Easy'):
    '''
    List all the trajectory folders, which is named as 'P0XX'
    '''
    assert(easy_hard=='Easy' or easy_hard=='Hard')
    traj_gen = container_client.walk_blobs(name_starts_with=envname + '/' + easy_hard+'/')
    trajlist = []
    for traj in traj_gen:
        trajname = traj.name
        trajname_split = trajname.split('/')
        trajname_split = [tt for tt in trajname_split if len(tt)>0]
        if trajname_split[-1][0] == 'P':
            trajlist.append(trajname)
    return trajlist

def _list_blobs_in_folder(folder_name):
    """
    List all blobs in a virtual folder in an Azure blob container
    """
    
    files = []
    generator = container_client.list_blobs(name_starts_with=folder_name)
    for blob in generator:
        files.append(blob.name)
    return files

def get_image_list(trajdir, left_right = 'left'):
    assert(left_right == 'left' or left_right == 'right')
    files = _list_blobs_in_folder(trajdir + '/image_' + left_right + '/')
    files = [fn for fn in files if fn.endswith('.png')]
    return files

def get_depth_list(trajdir, left_right = 'left'):
    assert(left_right == 'left' or left_right == 'right')
    files = _list_blobs_in_folder(trajdir + '/depth_' + left_right + '/')
    files = [fn for fn in files if fn.endswith('.npy')]
    return files

def get_flow_list(trajdir, ):
    files = _list_blobs_in_folder(trajdir + '/flow/')
    files = [fn for fn in files if fn.endswith('flow.npy')]
    return files

def get_flow_mask_list(trajdir, ):
    files = _list_blobs_in_folder(trajdir + '/flow/')
    files = [fn for fn in files if fn.endswith('mask.npy')]
    return files

def get_posefile(trajdir, left_right = 'left'):
    assert(left_right == 'left' or left_right == 'right')
    return trajdir + '/pose_' + left_right + '.txt'

def get_seg_list(trajdir, left_right = 'left'):
    assert(left_right == 'left' or left_right == 'right')
    files = _list_blobs_in_folder(trajdir + '/seg_' + left_right + '/')
    files = [fn for fn in files if fn.endswith('.npy')]
    return files

列出環境

envlist = get_environment_list()
print('Find {} environments..'.format(len(envlist)))
print(envlist)

列出第一個環境中的「簡單」軌跡

diff_level = 'Easy'
env_ind = 0
trajlist = get_trajectory_list(envlist[env_ind], easy_hard = diff_level)
print('Find {} trajectories in {}'.format(len(trajlist), envlist[env_ind]+diff_level))
print(trajlist)

列出一個軌跡中的所有資料檔案

traj_ind = 1
traj_dir = trajlist[traj_ind]

left_img_list = get_image_list(traj_dir, left_right = 'left')
print('Find {} left images in {}'.format(len(left_img_list), traj_dir))  

right_img_list = get_image_list(traj_dir, left_right = 'right')
print('Find {} right images in {}'.format(len(right_img_list), traj_dir))

left_depth_list = get_depth_list(traj_dir, left_right = 'left')
print('Find {} left depth files in {}'.format(len(left_depth_list), traj_dir))

right_depth_list = get_depth_list(traj_dir, left_right = 'right')
print('Find {} right depth files in {}'.format(len(right_depth_list), traj_dir))

left_seg_list = get_seg_list(traj_dir, left_right = 'left')
print('Find {} left segmentation files in {}'.format(len(left_seg_list), traj_dir))

right_seg_list = get_seg_list(traj_dir, left_right = 'left')
print('Find {} right segmentation files in {}'.format(len(right_seg_list), traj_dir))

flow_list = get_flow_list(traj_dir)
print('Find {} flow files in {}'.format(len(flow_list), traj_dir)) 

flow_mask_list = get_flow_mask_list(traj_dir)
print('Find {} flow mask files in {}'.format(len(flow_mask_list), traj_dir)) 

left_pose_file = get_posefile(traj_dir, left_right = 'left')
print('Left pose file: {}'.format(left_pose_file))

right_pose_file = get_posefile(traj_dir, left_right = 'right')
print('Right pose file: {}'.format(right_pose_file))

資料下載函式

def read_numpy_file(numpy_file,):
    '''
    return a numpy array given the file path
    '''
    bc = container_client.get_blob_client(blob=numpy_file)
    data = bc.download_blob()
    ee = io.BytesIO(data.content_as_bytes())
    ff = np.load(ee)
    return ff


def read_image_file(image_file,):
    '''
    return a uint8 numpy array given the file path  
    '''
    bc = container_client.get_blob_client(blob=image_file)
    data = bc.download_blob()
    ee = io.BytesIO(data.content_as_bytes())
    img=cv2.imdecode(np.asarray(bytearray(ee.read()),dtype=np.uint8), cv2.IMREAD_COLOR)
    im_rgb = img[:, :, [2, 1, 0]] # BGR2RGB
    return im_rgb

資料視覺效果函式

def depth2vis(depth, maxthresh = 50):
    depthvis = np.clip(depth,0,maxthresh)
    depthvis = depthvis/maxthresh*255
    depthvis = depthvis.astype(np.uint8)
    depthvis = np.tile(depthvis.reshape(depthvis.shape+(1,)), (1,1,3))

    return depthvis

def seg2vis(segnp):
    colors = [(205, 92, 92), (0, 255, 0), (199, 21, 133), (32, 178, 170), (233, 150, 122), (0, 0, 255), (128, 0, 0), (255, 0, 0), (255, 0, 255), (176, 196, 222), (139, 0, 139), (102, 205, 170), (128, 0, 128), (0, 255, 255), (0, 255, 255), (127, 255, 212), (222, 184, 135), (128, 128, 0), (255, 99, 71), (0, 128, 0), (218, 165, 32), (100, 149, 237), (30, 144, 255), (255, 0, 255), (112, 128, 144), (72, 61, 139), (165, 42, 42), (0, 128, 128), (255, 255, 0), (255, 182, 193), (107, 142, 35), (0, 0, 128), (135, 206, 235), (128, 0, 0), (0, 0, 255), (160, 82, 45), (0, 128, 128), (128, 128, 0), (25, 25, 112), (255, 215, 0), (154, 205, 50), (205, 133, 63), (255, 140, 0), (220, 20, 60), (255, 20, 147), (95, 158, 160), (138, 43, 226), (127, 255, 0), (123, 104, 238), (255, 160, 122), (92, 205, 92),]
    segvis = np.zeros(segnp.shape+(3,), dtype=np.uint8)

    for k in range(256):
        mask = segnp==k
        colorind = k % len(colors)
        if np.sum(mask)>0:
            segvis[mask,:] = colors[colorind]

    return segvis

def _calculate_angle_distance_from_du_dv(du, dv, flagDegree=False):
    a = np.arctan2( dv, du )

    angleShift = np.pi

    if ( True == flagDegree ):
        a = a / np.pi * 180
        angleShift = 180
        # print("Convert angle from radian to degree as demanded by the input file.")

    d = np.sqrt( du * du + dv * dv )

    return a, d, angleShift

def flow2vis(flownp, maxF=500.0, n=8, mask=None, hueMax=179, angShift=0.0): 
    """
    Show a optical flow field as the KITTI dataset does.
    Some parts of this function is the transform of the original MATLAB code flow_to_color.m.
    """

    ang, mag, _ = _calculate_angle_distance_from_du_dv( flownp[:, :, 0], flownp[:, :, 1], flagDegree=False )

    # Use Hue, Saturation, Value colour model 
    hsv = np.zeros( ( ang.shape[0], ang.shape[1], 3 ) , dtype=np.float32)

    am = ang < 0
    ang[am] = ang[am] + np.pi * 2

    hsv[ :, :, 0 ] = np.remainder( ( ang + angShift ) / (2*np.pi), 1 )
    hsv[ :, :, 1 ] = mag / maxF * n
    hsv[ :, :, 2 ] = (n - hsv[:, :, 1])/n

    hsv[:, :, 0] = np.clip( hsv[:, :, 0], 0, 1 ) * hueMax
    hsv[:, :, 1:3] = np.clip( hsv[:, :, 1:3], 0, 1 ) * 255
    hsv = hsv.astype(np.uint8)

    rgb = cv2.cvtColor(hsv, cv2.COLOR_HSV2RGB)

    if ( mask is not None ):
        mask = mask > 0
        rgb[mask] = np.array([0, 0 ,0], dtype=np.uint8)

    return rgb

下載並視覺化

data_ind = 173 # randomly select one frame (data_ind < TRAJ_LEN)

left_img = read_image_file(left_img_list[data_ind])
right_img = read_image_file(right_img_list[data_ind])

# Visualize the left and right RGB images
plt.figure(figsize=(12, 5))
plt.subplot(121)
plt.imshow(left_img)
plt.title('Left Image')
plt.subplot(122)
plt.imshow(right_img)
plt.title('Right Image')
plt.show()

# Visualize the left and right depth files
left_depth = read_numpy_file(left_depth_list[data_ind])
left_depth_vis = depth2vis(left_depth)

right_depth = read_numpy_file(right_depth_list[data_ind])
right_depth_vis = depth2vis(right_depth)

plt.figure(figsize=(12, 5))
plt.subplot(121)
plt.imshow(left_depth_vis)
plt.title('Left Depth')
plt.subplot(122)
plt.imshow(right_depth_vis)
plt.title('Right Depth')
plt.show()

# Visualize the left and right segmentation files
left_seg = read_numpy_file(left_seg_list[data_ind])
left_seg_vis = seg2vis(left_seg)

right_seg = read_numpy_file(right_seg_list[data_ind])
right_seg_vis = seg2vis(right_seg)

plt.figure(figsize=(12, 5))
plt.subplot(121)
plt.imshow(left_seg_vis)
plt.title('Left Segmentation')
plt.subplot(122)
plt.imshow(right_seg_vis)
plt.title('Right Segmentation')
plt.show()

# Visualize the flow and mask files

flow = read_numpy_file(flow_list[data_ind])
flow_vis = flow2vis(flow)

flow_mask = read_numpy_file(flow_mask_list[data_ind])
flow_vis_w_mask = flow2vis(flow, mask = flow_mask)

plt.figure(figsize=(12, 5))
plt.subplot(121)
plt.imshow(flow_vis)
plt.title('Optical Flow')
plt.subplot(122)
plt.imshow(flow_vis_w_mask)
plt.title('Optical Flow w/ Mask')
plt.show()

下一步

檢視開放資料集目錄中的其餘資料集。

共用方式為