我们提供安全,免费的手游软件下载!

安卓手机游戏下载_安卓手机软件下载_安卓手机应用免费下载-先锋下载

当前位置: 主页 > 软件教程 > 软件教程

阅读mmdetection3d框架的源码探索其构建dataset的流程

来源:网络 更新时间:2024-06-18 15:30:26

在查看一些基于 mmdetection3d 构建的代码的时候,一开始会摸不着头脑,它的 dataset 到底是怎么构造的?

接下来就直接下载 mmdetection3d 这个仓库,然后去分析里面的代码。

可以看到config(比如 mmdetection3d/configs/centerpoint/centerpoint_02pillar_second_secfpn_4x8_cyclic_20e_nus.py )有这样的字段:

train_pipeline = [
    # 属于pipelines.loading中的操作
    dict(
        type='LoadPointsFromFile',
        coord_type='LIDAR',
        load_dim=5,
        use_dim=5,
        file_client_args=file_client_args),
    # 输出的东西已经是一个名为results的字典了,里面最重要的就是results['points']了,是LiDARPoints类
    dict(
        type='LoadPointsFromMultiSweeps',
        sweeps_num=9,
        use_dim=[0, 1, 2, 3, 4],
        file_client_args=file_client_args,
        pad_empty_sweeps=True,
        remove_close=True),
    # 接收上面输出的results字典,然后对于里面的results['points']做增密操作。
    # 输出的字典有15个key:dict_keys(['sample_idx', 'pts_filename', 'sweeps', 'timestamp', 'ann_info', 'img_fields', 'bbox3d_fields', 'pts_mask_fields', 'pts_seg_fields', 'bbox_fields', 'mask_fields', 'seg_fields', 'box_type_3d', 'box_mode_3d', 'points'])
    dict(type='LoadAnnotations3D', with_bbox_3d=True, with_label_3d=True),
    # 输出的字典有17个key:dict_keys(['sample_idx', 'pts_filename', 'sweeps', 'timestamp', 'ann_info', 'img_fields', 'bbox3d_fields', 'pts_mask_fields', 'pts_seg_fields', 'bbox_fields', 'mask_fields', 'seg_fields', 'box_type_3d', 'box_mode_3d', 'points', 'gt_bboxes_3d', 'gt_labels_3d'])
    # 所以主要增加的键来自于这俩方法:self._load_bboxes_3d()和self._load_labels_3d()
    
    # 属于piplines.transform_3d
    dict(type='ObjectSample', db_sampler=db_sampler),
    dict(
        type='GlobalRotScaleTrans',
        rot_range=[-0.3925, 0.3925],
        scale_ratio_range=[0.95, 1.05],
        translation_std=[0, 0, 0]),
    dict(
        type='RandomFlip3D',
        sync_2d=False,
        flip_ratio_bev_horizontal=0.5,
        flip_ratio_bev_vertical=0.5),
    dict(type='PointsRangeFilter', point_cloud_range=point_cloud_range),
    dict(type='ObjectRangeFilter', point_cloud_range=point_cloud_range),
    dict(type='ObjectNameFilter', classes=class_names),
    dict(type='PointShuffle'),

    # 属于piplines.formatting
    dict(type='DefaultFormatBundle3D', class_names=class_names),
    dict(type='Collect3D', keys=['points', 'gt_bboxes_3d', 'gt_labels_3d'])
    # 最后只留下'image_meta'和上面的这三个keys。
]

只能大概猜测它是通过这些预先定义好了的类去规范整个dataset的加载过程吗?但是具体又是体现在哪个地方呢?或者说实际上load的时候怎么取到这些字典里面的值的呢?

下面来跟着代码走一下,

运行 python tools/train.py configs/centerpoint/centerpoint_02pillar_second_secfpn_4x8_cyclic_20e_nus.py 然后打断点

首先看到 tools/train.py 中的这里

经过build_dataset方法得到的dataset。

那么就要进去看build_dataset是做了啥

因为config里面写的data的type是 type='CBGSDataset' ,所以会先走第39行,然后再次调用该函数,再走第43行。

然后就是进入 build_from_cfg 函数了,

它最重要的代码就是截图所示,本质上是:

  • args 中移除键 'type' 并获取其值 obj_type
  • 如果 obj_type 是字符串,从 registry 中查找对应的类 obj_cls
  • 如果 obj_type 是类,直接使用 obj_type 作为 obj_cls
  • 如果 obj_type 既不是字符串也不是类,抛出类型错误。
  • 尝试使用 args 中的参数来实例化 obj_cls
  • 如果实例化过程中发生异常,捕获异常并抛出带有类名的异常信息,以便调试。

那么在这一层的 build_from_cfg 函数中, obj_cls 指的是 ,也就是要构建这个类,然后返回去。

再往里走看这个 NuScenesDataset 类的构造函数,它的基类是 Custom3DDataset

而且这里面就有pipeline参数了,如果在图中它执行父类构造函数的时候停下来来看这个 pipeline ,可以看到是

pipeline: [{'type': 'LoadPointsFromFile', 'coord_type': 'LIDAR', 'load_dim': 5, 'use_dim': 5, 'file_client_args': {...}}, {'type': 'LoadPointsFromMultiSweeps', 'sweeps_num': 9, 'use_dim': [...], 'file_client_args': {...}, 'pad_empty_sweeps': True, 'remove_close': True}, {'type': 'LoadAnnotations3D', 'with_bbox_3d': True, 'with_label_3d': True}, {'type': 'ObjectSample', 'db_sampler': {...}}, {'type': 'GlobalRotScaleTrans', 'rot_range': [...], 'scale_ratio_range': [...], 'translation_std': [...]}, {'type': 'RandomFlip3D', 'sync_2d': False, 'flip_ratio_bev_horizontal': 0.5, 'flip_ratio_bev_vertical': 0.5}, {'type': 'PointsRangeFilter', 'point_cloud_range': [...]}, {'type': 'ObjectRangeFilter', 'point_cloud_range': [...]}, {'type': 'ObjectNameFilter', 'classes': [...]}, {'type': 'PointShuffle'}, {'type': 'DefaultFormatBundle3D', 'class_names': [...]}, {'type': 'Collect3D', 'keys': [...]}]

它其实就是上面config里面的东西。

然后就执行父类构造函数啦,把 pipeline 做了一下Compose

接着就可以看到Compose,大概是把这些流程都组合起来的意思,这个没有深究

所以,如果 transforms 列表中的元素 transform callable 的(即是类),那么就直接append到 self.transform 里面,反之则要再重复调用 build_from_cfg 函数,然后用 transform(是个字典) PIPELINES 去构建这个类。再append到 self.transform 里面。

至于这俩参数是什么,放一下详细解答:

transform是一个字典,例:

{'type': 'LoadPointsFromMultiSweeps', 'sweeps_num': 9, 'use_dim': [0, 1, 2, 3, 4], 'file_client_args': {'backend': 'disk'}, 'pad_empty_sweeps': True, 'remove_close': True}

PIPELINE是一个注册表,

Registry(name=pipeline, items={'Compose': , 'AutoAugment': , 'Shear': , 'Rotate': , 'Translate': , 'ColorTransform': , 'EqualizeTransform': , 'BrightnessTransform': , 'ContrastTransform': , 'ToTensor': , 'ImageToTensor': , 'Transpose': , 'ToDataContainer': , 'Collect': , 'WrapFieldsToLists': , 'InstaBoost': , 'LoadImageFromFile': , 'LoadImageFromWebcam': , 'LoadMultiChannelImageFromFiles': , 'LoadAnnotations': , 'LoadPanopticAnnotations': , 'LoadProposals': , 'FilterAnnotations': , 'MultiScaleFlipAug': , 'Resize': , 'RandomFlip': , 'RandomShift': , 'Pad': , 'Normalize': , 'RandomCrop': , 'SegRescale': , 'PhotoMetricDistortion': , 'Expand': , 'MinIoURandomCrop': , 'Corrupt': , 'Albu': , 'RandomCenterCropPad': , 'CutOut': , 'Mosaic': , 'MixUp': , 'RandomAffine': , 'YOLOXHSVRandomAug': , 'CopyPaste': , 'DefaultFormatBundle': , 'Collect3D': , 'DefaultFormatBundle3D': , 'LoadMultiViewImageFromFiles': , 'LoadImageFromFileMono3D': , 'LoadPointsFromMultiSweeps': , 'PointSegClassMapping': , 'NormalizePointsColor': , 'LoadPointsFromFile': , 'LoadPointsFromDict': , 'LoadAnnotations3D': , 'MultiScaleFlipAug3D': , 'RandomDropPointsColor': , 'RandomFlip3D': , 'RandomJitterPoints': , 'ObjectSample': , 'ObjectNoise': , 'GlobalAlignment': , 'GlobalRotScaleTrans': , 'PointShuffle': , 'ObjectRangeFilter': , 'PointsRangeFilter': , 'ObjectNameFilter': , 'PointSample': , 'IndoorPointSample': , 'IndoorPatchPointSample': , 'BackgroundPointsFilter': , 'VoxelBasedPointSampler': , 'AffineResize': , 'RandomShiftScale': })

那么最后 self.transforms 是什么,它是一个长度为12的列表,装的其实就是上面config里面的那一堆用预定义好的类来表示的操作:

再看到 Compose 类的 __call__ 方法,

可以看到它其实是将data连续地走过了 self.transforms 里面的所有类,也就是这样来完成 pipeline 的具体实现的。

然后因为整个 train.py 中的dataset是 CBGSDataset ,所以去看它的 __getitem__()

发现相当于调它的 self.dataset __getitem__()

而前面有说过,它是包裹着 NuscenesDataset 的,即 self.dataset 的类就是 NuscenesDataset

所以就得去看 NuscenesDataset __getitem__() 的时候返回的是什么?发现得去看它的父类 Custom3DDataset __getitem__()

点进去看 prepare_train_data()

其实就是取出了当前sample的图或者LiDAR的数据 input_dict

{input_dict: 'sample_idx': '0468b5e2192f40d88a4e3353c83d28a1', 'pts_filename': './data/nuscenes/samples/LIDAR_TOP/n008-2018-08-29-16-04-13-0400__LIDAR_TOP__1535573055397947.pcd.bin', 'sweeps': [{...}, {...}, {...}, {...}, {...}, {...}, {...}, {...}, {...}, {...}], 'timestamp': 1535573055.397947, 'ann_info': {'gt_bboxes_3d': LiDARInstance3DBoxes(
    tensor([[-1.3053e+01,  2.3971e+01, -8.4190e-01,  1.9640e+00,  4.8260e+00,
          1.7050e+00, -1.5572e+00,  5.6951e-05,  9.9719e-04],
        [ 4.0986e+00, -5.1065e+01, -3.7889e+00,  2.2280e+00,  4.9170e+00,
          2.7230e+00, -1.5088e+00,  0.0000e+00,  0.0000e+00],
        [ 5.9061e+00, -2.5881e+01, -2.8135e+00,  1.8990e+00,  4.5950e+00,
          1.6550e+00, -1.5894e+00,  1.4087e-02,  1.1929e-03],
        [-6.5956e-01, -1.9181e+01, -2.5623e+00,  2.0340e+00,  5.0230e+00,
          1.6550e+00, -3.1338e+00,  1.5844e-01,  8.3847e+00],
        [-1.3778e+01,  4.6664e+00, -1.7571e+00,  2.1550e+00,  5.5660e+00,
          1.9950e+00, -4.6786e+00,  0.0000e+00,  0.0000e+00],
        [-1.4067e+01,  1.6442e+00, -1.7979e+00,  1.9260e+00,  4.8000e+00,
          1.6530e+00, -1.5948e+00,  2.7373e-02,  2.3441e-02],
        [-2.7166e+01, -3.5422e+00, -2.7281e+00,  1.7740e+00,  4.3200e+00,
          1.6430e+00, -1.5752e+00,  0.0000e+00,  0.0000e+00],
        [-1.5801e+01, -4.8262e+01, -3.9795e+00,  2.1070e+00,  4.9980e+00,
          1.8140e+00, -1.5487e+00, -6.7043e-02,  3.5860e-02],
        [-1.1155e+01,  7.9349e+01,  1.8333e+00,  2.2070e+00,  5.7040e+00,
          2.0920e+00, -4.6931e+00, -9.7979e-02,  2.6185e-03],
        [ 1.1782e+01, -2.5557e+00, -1.6983e+00,  1.8120e+00,  4.7570e+00,
          1.5840e+00, -4.6788e+00, -4.9902e-03,  2.8623e-04],
        [ 2.4727e+01, -1.3353e+01, -1.7999e+00,  1.9700e+00,  5.1180e+00,
          1.8120e+00, -1.5493e+00,  0.0000e+00,  0.0000e+00],
        [ 5.1789e+01,  1.9438e+01, -2.9660e-01,  1.9310e+00,  4.2030e+00,
          1.9810e+00, -2.5001e+00, -3.7849e-02, -2.4194e-01],
        [-9.9715e+00,  1.3736e+00, -1.6758e+00,  1.1540e+00,  7.9300e-01,
          1.7680e+00, -1.2722e+00, -7.5932e-02, -4.1784e-01],
        [-3.4043e+01,  1.4688e+01, -7.7760e-01,  8.2200e-01,  9.3900e-01,
          2.0000e+00, -3.1572e+00, -1.7419e-02,  1.8329e+00],
        [-9.1596e+00,  6.2475e+01,  1.2546e+00,  4.9400e-01,  6.4900e-01,
          1.8350e+00, -3.2893e+00,  3.7277e-02,  2.1861e-02],
        [-1.3011e+01,  3.4884e+01, -5.2392e-01,  2.1080e+00,  4.9040e+00,
          1.6120e+00, -1.4880e+00, -2.8475e-04, -4.9859e-03],
        [ 5.6555e+00, -3.4311e+01, -3.3156e+00,  1.8730e+00,  4.4170e+00,
          1.8000e+00, -1.5213e+00, -1.0550e-03, -9.3994e-04],
        [-1.4270e+01, -7.5958e-01, -1.8951e+00,  1.9500e+00,  4.7310e+00,
          1.6730e+00, -1.5787e+00,  2.1642e-02,  1.0764e-02],
        [-2.5297e+01,  3.7611e+01, -8.4330e-01,  2.9440e+00,  4.1020e+00,
          1.5660e+00,  1.5101e+00, -8.7569e-03,  4.4248e-01],
        [-1.1432e+01, -2.0810e+01, -2.7025e+00,  6.1800e-01,  6.9100e-01,
          1.8900e+00, -1.7533e+00,  1.3554e+00,  3.2644e-01],
        [-2.6757e+01, -8.8425e+00, -2.7591e+00,  2.1220e+00,  4.5250e+00,
          1.8830e+00, -4.6994e+00,  2.8475e-04,  4.9859e-03],
        [ 5.8472e+00, -3.1766e+01, -3.3697e+00,  1.8280e+00,  4.3640e+00,
          2.0000e+00, -1.5213e+00,  2.8503e-03,  1.4843e-02],
        [-1.3683e+01,  1.8342e+01, -1.3403e+00,  2.0270e+00,  4.3520e+00,
          1.5400e+00, -1.6096e+00,  0.0000e+00,  0.0000e+00],
        [ 8.5600e+00, -1.8746e+01, -2.4114e+00,  6.7100e-01,  1.4200e+00,
          1.2430e+00, -3.7172e+00, -1.4976e-01, -9.7458e-02],
        [-1.3618e+01,  7.3931e+00, -1.4792e+00,  2.2650e+00,  5.4260e+00,
          1.8200e+00, -4.7102e+00, -1.8709e-03, -1.3796e-01],
        [ 3.5973e+01,  4.1933e+01,  4.0528e-01,  1.5150e+00,  3.4580e+00,
          2.0000e+00, -3.1788e+00, -8.3159e-03,  2.2258e-01],
        [ 4.9209e+01,  2.1194e+01, -2.8792e-01,  1.8320e+00,  3.9100e+00,
          2.0360e+00, -2.6253e+00,  5.8494e-02, -1.1541e-01],
        [-2.6333e+01,  4.6259e+00, -2.2232e+00,  1.9040e+00,  3.9660e+00,
          1.3920e+00, -4.6994e+00,  0.0000e+00,  0.0000e+00],
        [-1.3432e+01,  1.2661e+01, -1.3381e+00,  2.0670e+00,  4.9390e+00,
          1.3830e+00, -1.5171e+00,  1.7713e-02, -5.8042e-02],
        [-2.4358e+01,  7.7235e+01,  6.3165e-01,  2.0340e+00,  4.0860e+00,
          1.5930e+00, -4.7057e+00,  0.0000e+00,  0.0000e+00],
        [-2.7353e+01, -5.3840e+01, -4.7062e+00,  2.3260e+00,  5.0220e+00,
          2.6770e+00,  1.5664e+00,  0.0000e+00,  0.0000e+00],
        [-1.8183e+01, -1.5049e+01, -2.6358e+00,  5.1600e-01,  6.8800e-01,
          1.8210e+00, -4.5943e+00, -4.5360e-01,  7.0038e-02],
        [ 5.7435e+00, -3.6962e+01, -3.4783e+00,  1.9290e+00,  4.5690e+00,
          1.8810e+00, -1.5213e+00, -3.9922e-03,  2.2898e-04],
        [ 1.2472e+01,  1.7255e+01, -8.0609e-01,  2.0270e+00,  4.5710e+00,
          1.7280e+00, -1.5121e+00, -1.9961e-03,  1.1449e-04],
        [-3.4475e+01,  3.4049e+01,  1.8847e-02,  5.1600e-01,  7.1900e-01,
          1.8210e+00, -3.1083e+00,  2.2408e-02,  1.3386e+00]])), 'gt_labels_3d': array([0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 4, 8, 8, 8, 0, 0, 0, 0, 8, 0, 0,
       0, 7, 1, 2, 4, 0, 0, 0, 0, 8, 0, 0, 8]), 'gt_names': array(['car', 'truck', 'car', 'car', 'truck', 'car', 'car', 'car', 'car',
       'car', 'car', 'trailer', 'pedestrian', 'pedestrian', 'pedestrian',
       'car', 'car', 'car', 'car', 'pedestrian', 'car', 'car', 'car',
       'bicycle', 'truck', 'construction_vehicle', 'trailer', 'car',
       'car', 'car', 'car', 'pedestrian', 'car', 'car', 'pedestrian'],
      dtype='

然后再经过一个 self.prepipeline 之后变成:

input_dict: {'sample_idx': '0468b5e2192f40d88a4e3353c83d28a1', 'pts_filename': './data/nuscenes/samples/LIDAR_TOP/n008-2018-08-29-16-04-13-0400__LIDAR_TOP__1535573055397947.pcd.bin', 'sweeps': [{...}, {...}, {...}, {...}, {...}, {...}, {...}, {...}, {...}, {...}], 'timestamp': 1535573055.397947, 'ann_info': {'gt_bboxes_3d': LiDARInstance3DBoxes(
    tensor([[-1.3053e+01,  2.3971e+01, -8.4190e-01,  1.9640e+00,  4.8260e+00,
          1.7050e+00, -1.5572e+00,  5.6951e-05,  9.9719e-04],
        [ 4.0986e+00, -5.1065e+01, -3.7889e+00,  2.2280e+00,  4.9170e+00,
          2.7230e+00, -1.5088e+00,  0.0000e+00,  0.0000e+00],
        [ 5.9061e+00, -2.5881e+01, -2.8135e+00,  1.8990e+00,  4.5950e+00,
          1.6550e+00, -1.5894e+00,  1.4087e-02,  1.1929e-03],
        [-6.5956e-01, -1.9181e+01, -2.5623e+00,  2.0340e+00,  5.0230e+00,
          1.6550e+00, -3.1338e+00,  1.5844e-01,  8.3847e+00],
        [-1.3778e+01,  4.6664e+00, -1.7571e+00,  2.1550e+00,  5.5660e+00,
          1.9950e+00, -4.6786e+00,  0.0000e+00,  0.0000e+00],
        [-1.4067e+01,  1.6442e+00, -1.7979e+00,  1.9260e+00,  4.8000e+00,
          1.6530e+00, -1.5948e+00,  2.7373e-02,  2.3441e-02],
        [-2.7166e+01, -3.5422e+00, -2.7281e+00,  1.7740e+00,  4.3200e+00,
          1.6430e+00, -1.5752e+00,  0.0000e+00,  0.0000e+00],
        [-1.5801e+01, -4.8262e+01, -3.9795e+00,  2.1070e+00,  4.9980e+00,
          1.8140e+00, -1.5487e+00, -6.7043e-02,  3.5860e-02],
        [-1.1155e+01,  7.9349e+01,  1.8333e+00,  2.2070e+00,  5.7040e+00,
          2.0920e+00, -4.6931e+00, -9.7979e-02,  2.6185e-03],
        [ 1.1782e+01, -2.5557e+00, -1.6983e+00,  1.8120e+00,  4.7570e+00,
          1.5840e+00, -4.6788e+00, -4.9902e-03,  2.8623e-04],
        [ 2.4727e+01, -1.3353e+01, -1.7999e+00,  1.9700e+00,  5.1180e+00,
          1.8120e+00, -1.5493e+00,  0.0000e+00,  0.0000e+00],
        [ 5.1789e+01,  1.9438e+01, -2.9660e-01,  1.9310e+00,  4.2030e+00,
          1.9810e+00, -2.5001e+00, -3.7849e-02, -2.4194e-01],
        [-9.9715e+00,  1.3736e+00, -1.6758e+00,  1.1540e+00,  7.9300e-01,
          1.7680e+00, -1.2722e+00, -7.5932e-02, -4.1784e-01],
        [-3.4043e+01,  1.4688e+01, -7.7760e-01,  8.2200e-01,  9.3900e-01,
          2.0000e+00, -3.1572e+00, -1.7419e-02,  1.8329e+00],
        [-9.1596e+00,  6.2475e+01,  1.2546e+00,  4.9400e-01,  6.4900e-01,
          1.8350e+00, -3.2893e+00,  3.7277e-02,  2.1861e-02],
        [-1.3011e+01,  3.4884e+01, -5.2392e-01,  2.1080e+00,  4.9040e+00,
          1.6120e+00, -1.4880e+00, -2.8475e-04, -4.9859e-03],
        [ 5.6555e+00, -3.4311e+01, -3.3156e+00,  1.8730e+00,  4.4170e+00,
          1.8000e+00, -1.5213e+00, -1.0550e-03, -9.3994e-04],
        [-1.4270e+01, -7.5958e-01, -1.8951e+00,  1.9500e+00,  4.7310e+00,
          1.6730e+00, -1.5787e+00,  2.1642e-02,  1.0764e-02],
        [-2.5297e+01,  3.7611e+01, -8.4330e-01,  2.9440e+00,  4.1020e+00,
          1.5660e+00,  1.5101e+00, -8.7569e-03,  4.4248e-01],
        [-1.1432e+01, -2.0810e+01, -2.7025e+00,  6.1800e-01,  6.9100e-01,
          1.8900e+00, -1.7533e+00,  1.3554e+00,  3.2644e-01],
        [-2.6757e+01, -8.8425e+00, -2.7591e+00,  2.1220e+00,  4.5250e+00,
          1.8830e+00, -4.6994e+00,  2.8475e-04,  4.9859e-03],
        [ 5.8472e+00, -3.1766e+01, -3.3697e+00,  1.8280e+00,  4.3640e+00,
          2.0000e+00, -1.5213e+00,  2.8503e-03,  1.4843e-02],
        [-1.3683e+01,  1.8342e+01, -1.3403e+00,  2.0270e+00,  4.3520e+00,
          1.5400e+00, -1.6096e+00,  0.0000e+00,  0.0000e+00],
        [ 8.5600e+00, -1.8746e+01, -2.4114e+00,  6.7100e-01,  1.4200e+00,
          1.2430e+00, -3.7172e+00, -1.4976e-01, -9.7458e-02],
        [-1.3618e+01,  7.3931e+00, -1.4792e+00,  2.2650e+00,  5.4260e+00,
          1.8200e+00, -4.7102e+00, -1.8709e-03, -1.3796e-01],
        [ 3.5973e+01,  4.1933e+01,  4.0528e-01,  1.5150e+00,  3.4580e+00,
          2.0000e+00, -3.1788e+00, -8.3159e-03,  2.2258e-01],
        [ 4.9209e+01,  2.1194e+01, -2.8792e-01,  1.8320e+00,  3.9100e+00,
          2.0360e+00, -2.6253e+00,  5.8494e-02, -1.1541e-01],
        [-2.6333e+01,  4.6259e+00, -2.2232e+00,  1.9040e+00,  3.9660e+00,
          1.3920e+00, -4.6994e+00,  0.0000e+00,  0.0000e+00],
        [-1.3432e+01,  1.2661e+01, -1.3381e+00,  2.0670e+00,  4.9390e+00,
          1.3830e+00, -1.5171e+00,  1.7713e-02, -5.8042e-02],
        [-2.4358e+01,  7.7235e+01,  6.3165e-01,  2.0340e+00,  4.0860e+00,
          1.5930e+00, -4.7057e+00,  0.0000e+00,  0.0000e+00],
        [-2.7353e+01, -5.3840e+01, -4.7062e+00,  2.3260e+00,  5.0220e+00,
          2.6770e+00,  1.5664e+00,  0.0000e+00,  0.0000e+00],
        [-1.8183e+01, -1.5049e+01, -2.6358e+00,  5.1600e-01,  6.8800e-01,
          1.8210e+00, -4.5943e+00, -4.5360e-01,  7.0038e-02],
        [ 5.7435e+00, -3.6962e+01, -3.4783e+00,  1.9290e+00,  4.5690e+00,
          1.8810e+00, -1.5213e+00, -3.9922e-03,  2.2898e-04],
        [ 1.2472e+01,  1.7255e+01, -8.0609e-01,  2.0270e+00,  4.5710e+00,
          1.7280e+00, -1.5121e+00, -1.9961e-03,  1.1449e-04],
        [-3.4475e+01,  3.4049e+01,  1.8847e-02,  5.1600e-01,  7.1900e-01,
          1.8210e+00, -3.1083e+00,  2.2408e-02,  1.3386e+00]])), 'gt_labels_3d': array([0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 4, 8, 8, 8, 0, 0, 0, 0, 8, 0, 0,
       0, 7, 1, 2, 4, 0, 0, 0, 0, 8, 0, 0, 8]), 'gt_names': array(['car', 'truck', 'car', 'car', 'truck', 'car', 'car', 'car', 'car',
       'car', 'car', 'trailer', 'pedestrian', 'pedestrian', 'pedestrian',
       'car', 'car', 'car', 'car', 'pedestrian', 'car', 'car', 'car',
       'bicycle', 'truck', 'construction_vehicle', 'trailer', 'car',
       'car', 'car', 'car', 'pedestrian', 'car', 'car', 'pedestrian'],
      dtype=', 'box_mode_3d': , 'points': DataContainer(tensor([[-1.2596e+01, -2.1491e+01, -1.2492e+00,  1.1000e+01,  4.9989e-01],
        [ 8.3029e+00, -5.2680e+00, -1.1252e+00,  3.0000e+00,  1.0011e-01],
        [-1.4655e+01,  8.3101e+00, -1.1988e+00,  4.0000e+00,  0.0000e+00],
        ...,
        [-1.0501e+01,  5.2707e+00, -8.7517e-01,  2.0000e+00,  4.9784e-02],
        [ 1.9245e+00, -2.5313e-01, -1.8701e+00,  9.0000e+00,  4.5012e-01],
        [ 7.9341e-03, -8.9713e-02, -3.8146e-03,  6.0000e+00,  0.0000e+00]])), 'gt_bboxes_3d': DataContainer(LiDARInstance3DBoxes(
    tensor([[ 1.1081e+01,  2.5450e+01, -8.5619e-01,  1.9973e+00,  4.9079e+00,
          1.7339e+00, -1.4964e+00, -1.4681e-04,  1.0051e-03],
        [-3.6702e+00, -2.6746e+01, -2.8613e+00,  1.9312e+00,  4.6730e+00,
          1.6831e+00, -1.4643e+00, -1.4377e-02, -5.0396e-05],
        [ 2.3823e+00, -1.9372e+01, -2.6058e+00,  2.0685e+00,  5.1083e+00,
          1.6831e+00,  8.0225e-02, -9.0984e-01,  8.4799e+00],
        [ 1.3541e+01,  5.9586e+00, -1.7869e+00,  2.1916e+00,  5.6605e+00,
          2.0289e+00,  1.6250e+00, -0.0000e+00,  0.0000e+00],
        [ 1.4103e+01,  2.9228e+00, -1.8284e+00,  1.9587e+00,  4.8815e+00,
          1.6811e+00, -1.4588e+00, -2.9825e-02,  2.1301e-02],
        [ 2.7837e+01, -1.1606e+00, -2.7744e+00,  1.8041e+00,  4.3933e+00,
          1.6709e+00, -1.4784e+00, -0.0000e+00,  0.0000e+00],
        [ 2.0320e+01, -4.7479e+01, -4.0470e+00,  2.1428e+00,  5.0828e+00,
          1.8448e+00, -1.5049e+00,  6.4712e-02,  4.2319e-02],
        [-1.1707e+01, -3.6420e+00, -1.7271e+00,  1.8428e+00,  4.8377e+00,
          1.6109e+00,  1.6252e+00,  5.0297e-03,  7.3593e-04],
        [-2.3856e+01, -1.5737e+01, -1.8305e+00,  2.0034e+00,  5.2049e+00,
          1.8428e+00, -1.5043e+00, -0.0000e+00,  0.0000e+00],
        [ 9.9788e+00,  2.2827e+00, -1.7043e+00,  1.1736e+00,  8.0646e-01,
          1.7980e+00, -1.7814e+00,  1.1426e-01, -4.1650e-01],
        [ 3.3174e+01,  1.7922e+01, -7.9080e-01,  8.3595e-01,  9.5494e-01,
          2.0339e+00,  1.0356e-01, -1.4616e-01,  1.8584e+00],
        [ 1.0063e+01,  3.6501e+01, -5.3281e-01,  2.1438e+00,  4.9872e+00,
          1.6394e+00, -1.5656e+00,  7.3405e-04, -5.0255e-03],
        [-2.6629e+00, -3.5264e+01, -3.3719e+00,  1.9048e+00,  4.4920e+00,
          1.8306e+00, -1.5323e+00,  1.1528e-03, -8.5792e-04],
        [ 1.4524e+01,  5.0582e-01, -1.9272e+00,  1.9831e+00,  4.8113e+00,
          1.7014e+00, -1.4749e+00, -2.2887e-02,  8.9703e-03],
        [ 2.2266e+01,  4.0362e+01, -8.5761e-01,  2.9940e+00,  4.1716e+00,
          1.5926e+00,  1.7194e+00, -3.0672e-02,  4.4903e-01],
        [ 1.3440e+01, -2.0060e+01, -2.7484e+00,  6.2849e-01,  7.0273e-01,
          1.9221e+00, -1.3003e+00, -1.4022e+00,  2.0957e-01],
        [ 2.7897e+01, -6.5666e+00, -2.8059e+00,  2.1580e+00,  4.6018e+00,
          1.9150e+00,  1.6458e+00, -7.3405e-04,  5.0255e-03],
        [-3.0846e+00, -3.2703e+01, -3.4269e+00,  1.8590e+00,  4.4381e+00,
          2.0339e+00, -1.5323e+00, -4.2140e-03,  1.4782e-02],
        [ 1.2222e+01,  1.9804e+01, -1.3630e+00,  2.0614e+00,  4.4259e+00,
          1.5661e+00, -1.4440e+00, -0.0000e+00,  0.0000e+00],
        [-6.9963e+00, -1.9755e+01, -2.4524e+00,  6.8239e-01,  1.4441e+00,
          1.2641e+00,  6.6358e-01,  1.6042e-01, -8.5346e-02],
        [ 1.3135e+01,  8.7066e+00, -1.5043e+00,  2.3034e+00,  5.5181e+00,
          1.8509e+00,  1.6566e+00,  1.4224e-02, -1.3959e-01],
        [-4.0190e+01,  3.9265e+01,  4.1216e-01,  1.5407e+00,  3.5167e+00,
          2.0339e+00,  1.2515e-01, -1.1467e-02,  2.2622e-01],
        [ 2.6263e+01,  7.0396e+00, -2.2609e+00,  1.9363e+00,  4.0333e+00,
          1.4156e+00,  1.6458e+00, -0.0000e+00,  0.0000e+00],
        [ 1.2475e+01,  1.4027e+01, -1.3608e+00,  2.1021e+00,  5.0228e+00,
          1.4065e+00, -1.5365e+00, -1.2757e-02, -6.0382e-02],
        [ 1.9765e+01, -1.3621e+01, -2.6806e+00,  5.2476e-01,  6.9968e-01,
          1.8519e+00,  1.5407e+00,  4.5326e-01,  1.1149e-01],
        [-2.5151e+00, -3.7958e+01, -3.5373e+00,  1.9617e+00,  4.6466e+00,
          1.9129e+00, -1.5323e+00,  4.0238e-03,  5.8874e-04],
        [-1.4177e+01,  1.6366e+01, -8.1977e-01,  2.0614e+00,  4.6486e+00,
          1.7573e+00, -1.5415e+00,  2.0119e-03,  2.9437e-04],
        [ 3.1882e+01,  3.7574e+01,  1.9167e-02,  5.2476e-01,  7.3120e-01,
          1.8519e+00,  5.4691e-02, -1.4233e-01,  1.3541e+00],
        [-1.5008e+01,  4.2524e+01,  5.4680e-01,  3.3987e+00,  6.4303e+00,
          2.9808e+00,  3.0873e+00, -1.2606e-01,  5.3561e-01],
        [ 1.7001e+00,  3.5514e+01,  6.0073e-01,  2.6858e+00,  7.2144e+00,
          2.1082e+00, -3.1104e+00, -0.0000e+00,  0.0000e+00],
        [-4.2222e+00, -6.2725e+00, -1.2759e+00,  2.1468e+00,  4.6262e+00,
          2.3868e+00,  3.0942e+00, -5.8198e-02, -1.3791e-02],
        [ 1.2403e+01, -8.7635e+00, -9.2305e-01,  2.5679e+00,  7.6843e+00,
          2.2902e+00,  1.6714e+00, -5.7650e-03,  8.0131e-02],
        [-2.5936e+01,  7.1599e+00,  3.5442e-01,  2.9574e+00,  1.1094e+01,
          4.5296e+00,  1.9989e+00, -0.0000e+00,  0.0000e+00],
        [-1.8544e+01,  3.2498e+01,  1.5000e+00,  3.1852e+00,  1.4053e+01,
          3.6672e+00,  1.3203e+00,  3.1848e+00,  1.4379e+00],
        [-2.5963e+01,  1.4695e+01,  1.3435e+00,  2.6004e+00,  7.5327e+00,
          2.8587e+00, -1.4322e+00, -9.8903e+00, -4.8384e-01],
        [-3.6054e+01,  1.4393e+01,  1.7633e+00,  2.6004e+00,  7.5327e+00,
          2.8587e+00, -1.4321e+00, -1.0276e+01, -8.6155e-02],
        [ 4.5645e+00, -1.6092e+01, -4.2814e-01,  2.3696e+00,  1.4248e+01,
          3.9550e+00,  1.1303e-01, -0.0000e+00,  0.0000e+00],
        [-8.5879e+00, -2.0463e+01, -2.0144e+00,  2.2119e+00,  6.5493e-01,
          1.2661e+00, -1.4716e+00, -0.0000e+00,  0.0000e+00],
        [-1.4564e+01, -7.3974e+00, -1.3176e+00,  3.0855e+00,  4.0170e-01,
          9.5596e-01,  1.2726e+00, -0.0000e+00,  0.0000e+00],
        [ 6.9650e+00, -1.0772e+01, -1.2148e+00,  5.2578e-01,  1.3058e+00,
          1.6607e+00, -2.5509e+00, -1.5342e+00, -3.7594e+00],
        [-3.3463e+01, -2.9462e+01, -2.1733e+00,  1.1888e+00,  2.0940e+00,
          1.9638e+00,  7.8985e-01,  4.2958e-02,  6.0744e-02],
        [ 9.4685e+00, -1.9686e+01, -9.4145e-01,  7.1392e-01,  2.1072e+00,
          1.6465e+00,  8.3660e-02,  6.2571e-03,  2.1583e-01],
        [ 1.9394e+01,  2.1123e+01, -1.1023e-02,  5.2578e-01,  1.3058e+00,
          1.6607e+00, -2.5509e+00, -1.7439e+00, -3.7068e+00],
        [-3.2073e+01,  1.0891e+01,  7.2284e-01,  5.3391e-01,  1.6495e+00,
          1.7248e+00, -1.3910e+00, -3.7946e+00,  9.3973e-02],
        [-9.1412e+00,  2.0717e+01,  7.1289e-01,  9.4884e-01,  2.1072e+00,
          1.8702e+00,  1.4224e+00,  4.5628e+00,  1.5586e+00],
        [-1.2042e+01,  1.3107e+01,  2.1672e-01,  5.3391e-01,  1.6495e+00,
          1.7248e+00, -1.5480e+00, -3.3165e+00, -4.4372e-01],
        [-8.6158e+00,  1.3458e+01,  7.0057e-02,  5.3391e-01,  1.6495e+00,
          1.7248e+00, -1.4433e+00, -3.5085e+00, -2.4859e-01],
        [ 5.0856e+00,  3.1870e+01,  1.0608e-01,  3.6916e-01,  3.6713e-01,
          7.2409e-01, -2.7731e+00,  3.2728e-02, -6.6736e-03],
        [-1.0092e+01, -6.3476e+00, -1.6635e+00,  5.1357e-01,  4.6272e-01,
          1.0526e+00, -1.5827e+00, -0.0000e+00,  0.0000e+00]]))), 'gt_labels_3d': DataContainer(tensor([0, 0, 0, 1, 0, 0, 0, 0, 0, 8, 8, 0, 0, 0, 0, 8, 0, 0, 0, 7, 1, 2, 0, 0,
        8, 0, 0, 8, 2, 2, 2, 2, 3, 3, 3, 3, 4, 5, 5, 6, 6, 6, 6, 7, 7, 7, 7, 9,
        9])), 'transformation_3d_flow': ['R', 'S', 'T', 'VF'], 'pcd_rotation': tensor([[ 0.9961, -0.0879,  0.0000],
        [ 0.0879,  0.9961,  0.0000],
        [ 0.0000,  0.0000,  1.0000]]), 'pcd_rotation_angle': -0.08799006818782418, 'pcd_scale_factor': 1.016974603680348, 'pcd_trans': array([0., 0., 0.]), 'flip': False, 'flip_direction': None, 'pcd_horizontal_flip': False, 'pcd_vertical_flip': True}

然后再将它经过那个pipeline,最后的example就是:

example 是一个只有这四个键的字典,

example['points'] DataContainer 类,而 DataContainer.data 才是 tensor 类。

那么 example['points'].data.shape: torch.Size([242193, 5])

所以到这里就知道它是怎么加载dataset的以及在训练的时候,每次 __getitem__ 的时候拿到的是什么样子的数据了。