[MMDetection3D] Pointpillars 학습

config file

voxel_size = [0.25, 0.25, 8]
model = dict(
    type='VoxelNet',
    voxel_layer=dict(
        max_num_points=64,
        point_cloud_range=[-50, -50, -5, 50, 50, 3],
        voxel_size=[0.25, 0.25, 8],
        max_voxels=(30000, 40000)),
    voxel_encoder=dict(
        type='HardVFE',
        in_channels=4,
        feat_channels=[64],
        with_distance=False,
        voxel_size=[0.25, 0.25, 8],
        point_cloud_range=[-50, -50, -5, 50, 50, 3],
        norm_cfg=dict(type='naiveSyncBN1d', eps=1e-3, momentum=0.01)),
    middle_encoder=dict(
        type='PointPillarsScatter', in_channels=64, output_shape=[400, 400]),
    backbone=dict(
        type='SECOND',
        in_channels=64,
        layer_nums=[3, 5, 5],
        layer_strides=[2, 2, 2],
        out_channels=[64, 128, 256]),
    neck=dict(
        type='SECONDFPN',
        in_channels=[64, 128, 256],
        upsample_strides=[1, 2, 4],
        out_channels=[128, 128, 128]),
    bbox_head=dict(
        type='Anchor3DHead',
        num_classes=1,
        in_channels=384,
        feat_channels=384,
        use_direction_classifier=True,
        assign_per_class=True,
        anchor_generator=dict(
            type='AlignedAnchor3DRangeGenerator',
            ranges=[[-50, -50, -5, 50, 50, 3]],
            sizes=[[3.9, 1.6, 1.56]],
            rotations=[0, 1.57],
            reshape_out=True),
        diff_rad_by_sin=True,
        bbox_coder=dict(type='DeltaXYZWLHRBBoxCoder'),
        loss_cls=dict(
            type='FocalLoss',
            use_sigmoid=True,
            gamma=2.0,
            alpha=0.25,
            loss_weight=1.0),
        loss_bbox=dict(
            type='SmoothL1Loss', beta=0.1111111111111111, loss_weight=2.0),
        loss_dir=dict(
            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.2)),
    train_cfg=dict(
        assigner=dict(
            type='MaxIoUAssigner',
            iou_calculator=dict(type='BboxOverlapsNearest3D'),
            pos_iou_thr=0.6,
            neg_iou_thr=0.45,
            min_pos_iou=0.45,
            ignore_iof_thr=-1),
        allowed_border=0,
        pos_weight=-1,
        debug=False),
    test_cfg=dict(
        use_rotate_nms=True,
        nms_across_levels=False,
        nms_thr=0.01,
        score_thr=0.1,
        min_bbox_size=0,
        nms_pre=100,
        max_num=50))
dataset_type = 'KittiDataset'
data_root = 'data/carla/'
class_names = ['vehicle']
point_cloud_range = [-50, -50, -5, 50, 50, 3]
input_modality = dict(use_lidar=True, use_camera=False)
file_client_args = dict(backend='disk')
db_sampler = dict(
    data_root='data/carla/',
    info_path='data/carla/kitti_dbinfos_train.pkl',
    rate=1.0,
    prepare=dict(
        filter_by_difficulty=[-1],
        filter_by_min_points=dict(Car=5, Pedestrian=5, Cyclist=5, vehicle=5)),
    classes=['vehicle'],
    sample_groups=dict(Car=15, Pedestrian=15, Cyclist=15, vehicle=15),
    points_loader=dict(
        type='LoadPointsFromFile',
        coord_type='LIDAR',
        load_dim=4,
        use_dim=4,
        file_client_args=dict(backend='disk')),
    file_client_args=dict(backend='disk'))
train_pipeline = [
    dict(type='LoadPointsFromFile', coord_type='LIDAR', load_dim=4, use_dim=4),
    dict(type='LoadAnnotations3D', with_bbox_3d=True, with_label_3d=True),
    dict(
        type='ObjectSample',
        db_sampler=dict(
            data_root='data/carla/',
            info_path='data/carla/kitti_dbinfos_train.pkl',
            rate=1.0,
            prepare=dict(
                filter_by_difficulty=[-1],
                filter_by_min_points=dict(vehicle=5)),
            sample_groups=dict(vehicle=15),
            classes=['vehicle']),
        use_ground_plane=True),
    dict(type='RandomFlip3D', flip_ratio_bev_horizontal=0.5),
    dict(
        type='GlobalRotScaleTrans',
        rot_range=[-0.78539816, 0.78539816],
        scale_ratio_range=[0.95, 1.05]),
    dict(
        type='PointsRangeFilter', point_cloud_range=[-50, -50, -5, 50, 50, 3]),
    dict(
        type='ObjectRangeFilter', point_cloud_range=[-50, -50, -5, 50, 50, 3]),
    dict(type='PointShuffle'),
    dict(type='DefaultFormatBundle3D', class_names=['vehicle']),
    dict(type='Collect3D', keys=['points', 'gt_bboxes_3d', 'gt_labels_3d'])
]
test_pipeline = [
    dict(type='LoadPointsFromFile', coord_type='LIDAR', load_dim=4, use_dim=4),
    dict(
        type='MultiScaleFlipAug3D',
        img_scale=(1333, 800),
        pts_scale_ratio=1,
        flip=False,
        transforms=[
            dict(
                type='GlobalRotScaleTrans',
                rot_range=[0, 0],
                scale_ratio_range=[1.0, 1.0],
                translation_std=[0, 0, 0]),
            dict(type='RandomFlip3D'),
            dict(
                type='PointsRangeFilter',
                point_cloud_range=[-40, -40, -3, 40, 40, 2]),
            dict(
                type='DefaultFormatBundle3D',
                class_names=['vehicle'],
                with_label=False),
            dict(type='Collect3D', keys=['points'])
        ])
]
eval_pipeline = [
    dict(
        type='LoadPointsFromFile',
        coord_type='LIDAR',
        load_dim=4,
        use_dim=4,
        file_client_args=dict(backend='disk')),
    dict(
        type='DefaultFormatBundle3D',
        class_names=['vehicle', 'walker', 'pet'],
        with_label=False),
    dict(type='Collect3D', keys=['points'])
]
data = dict(
    samples_per_gpu=1,
    workers_per_gpu=1,
    train=dict(
        type='RepeatDataset',
        times=2,
        dataset=dict(
            type='KittiDataset',
            data_root='data/carla/',
            ann_file='data/carla/kitti_infos_train.pkl',
            split='training',
            pts_prefix='velodyne',
            pipeline=[
                dict(
                    type='LoadPointsFromFile',
                    coord_type='LIDAR',
                    load_dim=4,
                    use_dim=4),
                dict(
                    type='LoadAnnotations3D',
                    with_bbox_3d=True,
                    with_label_3d=True),
                dict(
                    type='ObjectSample',
                    db_sampler=dict(
                        data_root='data/carla/',
                        info_path='data/carla/kitti_dbinfos_train.pkl',
                        rate=1.0,
                        prepare=dict(
                            filter_by_difficulty=[-1],
                            filter_by_min_points=dict(vehicle=5)),
                        sample_groups=dict(vehicle=15),
                        classes=['vehicle']),
                    use_ground_plane=True),
                dict(type='RandomFlip3D', flip_ratio_bev_horizontal=0.5),
                dict(
                    type='GlobalRotScaleTrans',
                    rot_range=[-0.78539816, 0.78539816],
                    scale_ratio_range=[0.95, 1.05]),
                dict(
                    type='PointsRangeFilter',
                    point_cloud_range=[-50, -50, -5, 50, 50, 3]),
                dict(
                    type='ObjectRangeFilter',
                    point_cloud_range=[-50, -50, -5, 50, 50, 3]),
                dict(type='PointShuffle'),
                dict(type='DefaultFormatBundle3D', class_names=['vehicle']),
                dict(
                    type='Collect3D',
                    keys=['points', 'gt_bboxes_3d', 'gt_labels_3d'])
            ],
            modality=dict(use_lidar=True, use_camera=False),
            classes=['vehicle'],
            test_mode=False,
            box_type_3d='LiDAR',
            file_client_args=dict(backend='disk'))),
    val=dict(
        type='KittiDataset',
        data_root='data/carla/',
        ann_file='data/carla/kitti_infos_val.pkl',
        split='training',
        pts_prefix='velodyne',
        pipeline=[
            dict(
                type='LoadPointsFromFile',
                coord_type='LIDAR',
                load_dim=4,
                use_dim=4),
            dict(
                type='MultiScaleFlipAug3D',
                img_scale=(1333, 800),
                pts_scale_ratio=1,
                flip=False,
                transforms=[
                    dict(
                        type='GlobalRotScaleTrans',
                        rot_range=[0, 0],
                        scale_ratio_range=[1.0, 1.0],
                        translation_std=[0, 0, 0]),
                    dict(type='RandomFlip3D'),
                    dict(
                        type='PointsRangeFilter',
                        point_cloud_range=[-50, -50, -5, 50, 50, 3]),
                    dict(
                        type='DefaultFormatBundle3D',
                        class_names=['vehicle'],
                        with_label=False),
                    dict(type='Collect3D', keys=['points'])
                ])
        ],
        modality=dict(use_lidar=True, use_camera=False),
        classes=['vehicle'],
        test_mode=True,
        box_type_3d='LiDAR',
        file_client_args=dict(backend='disk')),
    test=dict(
        type='KittiDataset',
        data_root='data/carla/',
        ann_file='data/carla/kitti_infos_val.pkl',
        split='training',
        pts_prefix='velodyne',
        pipeline=[
            dict(
                type='LoadPointsFromFile',
                coord_type='LIDAR',
                load_dim=4,
                use_dim=4),
            dict(
                type='MultiScaleFlipAug3D',
                img_scale=(1333, 800),
                pts_scale_ratio=1,
                flip=False,
                transforms=[
                    dict(
                        type='GlobalRotScaleTrans',
                        rot_range=[0, 0],
                        scale_ratio_range=[1.0, 1.0],
                        translation_std=[0, 0, 0]),
                    dict(type='RandomFlip3D'),
                    dict(
                        type='PointsRangeFilter',
                        point_cloud_range=[-50, -50, -5, 50, 50, 3]),
                    dict(
                        type='DefaultFormatBundle3D',
                        class_names=['vehicle'],
                        with_label=False),
                    dict(type='Collect3D', keys=['points'])
                ])
        ],
        modality=dict(use_lidar=True, use_camera=False),
        classes=['vehicle'],
        test_mode=True,
        box_type_3d='LiDAR',
        file_client_args=dict(backend='disk')))
evaluation = dict(
    interval=2,
    pipeline=[
        dict(
            type='LoadPointsFromFile',
            coord_type='LIDAR',
            load_dim=4,
            use_dim=4,
            file_client_args=dict(backend='disk')),
        dict(
            type='DefaultFormatBundle3D',
            class_names=['vehicle', 'walker', 'pet'],
            with_label=False),
        dict(type='Collect3D', keys=['points'])
    ])
lr = 0.001
optimizer = dict(type='AdamW', lr=0.001, betas=(0.95, 0.99), weight_decay=0.01)
optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
lr_config = dict(
    policy='cyclic',
    target_ratio=(10, 0.0001),
    cyclic_times=1,
    step_ratio_up=0.4)
momentum_config = dict(
    policy='cyclic',
    target_ratio=(0.8947368421052632, 1),
    cyclic_times=1,
    step_ratio_up=0.4)
runner = dict(type='EpochBasedRunner', max_epochs=120)
checkpoint_config = dict(interval=20)
log_config = dict(
    interval=10,
    hooks=[dict(type='TextLoggerHook'),
           dict(type='TensorboardLoggerHook')])
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = './work_dirs/hv_pointpillars_secfpn_1x1_160e_carla-3d-car'
load_from = None
resume_from = None
workflow = [('train', 1)]
opencv_num_threads = 0
mp_start_method = 'fork'
gpu_ids = [0]

 

pointcloud range와 voxel size를 수정해주었고

pcd를 velodyne_reduced가 아니라 velodyne로 하겠끔 수정해주었습니다.

 

학습 전 데이터 확인

python tools/misc/browse_dataset.py work_dirs/hv_pointpillars_secfpn_1x1_160e_carla-3d-car.py --task det --output-dir showdir/ --online

몇 개 확인해봤는데 잘 나옵니다.

 

 

Train

python ./tools/train.py work_dirs/hv_pointpillars_secfpn_1x1_160e_carla-3d-car.py --work-dir work_dirs

학습을 진행합니다. epochs는 config에서 확인할 수 있듯이 120번으로 해주었습니다.

 

학습이 잘 됩니다.

 

validation 항목은 안나왔는데 이건 코드를 따로 만들어줘야할 것 같아요.

 

Test

python tools/test.py ./work_dirs/hv_pointpillars_secfpn_1x1_160e_carla-3d-car.py work_dirs/latest.pth --eval 'mAP' --eval-options 'show=True' 'out_dir=./showdir'

 

 

{'KITTI/vehicle_3D_AP11_easy_strict': 0.0, 

'KITTI/vehicle_BEV_AP11_easy_strict': 0.0, 

'KITTI/vehicle_2D_AP11_easy_strict': 0.0,

'KITTI/vehicle_3D_AP11_moderate_strict': 0.0,

'KITTI/vehicle_BEV_AP11_moderate_strict': 0.0,

'KITTI/vehicle_2D_AP11_moderate_strict': 0.0,

'KITTI/vehicle_3D_AP11_hard_strict': 0.0,

'KITTI/vehicle_BEV_AP11_hard_strict': 0.0, '

KITTI/vehicle_2D_AP11_hard_strict': 0.0,

'KITTI/vehicle_3D_AP11_easy_loose': 0.0,

'KITTI/vehicle_BEV_AP11_easy_loose': 0.0,

'KITTI/vehicle_2D_AP11_easy_loose': 0.0,

'KITTI/vehicle_3D_AP11_moderate_loose': 0.0, '

KITTI/vehicle_BEV_AP11_moderate_loose': 0.0,

'KITTI/vehicle_2D_AP11_moderate_loose': 0.0,

'KITTI/vehicle_3D_AP11_hard_loose': 0.0,

'KITTI/vehicle_BEV_AP11_hard_loose': 0.0,

'KITTI/vehicle_2D_AP11_hard_loose': 0.0,

'KITTI/vehicle_3D_AP40_easy_strict': 0.0,

'KITTI/vehicle_BEV_AP40_easy_strict': 0.0,

'KITTI/vehicle_2D_AP40_easy_strict': 0.0,

'KITTI/vehicle_3D_AP40_moderate_strict': 0.0,

'KITTI/vehicle_BEV_AP40_moderate_strict': 0.0,

'KITTI/vehicle_2D_AP40_moderate_strict': 0.0,

'KITTI/vehicle_3D_AP40_hard_strict': 0.0,

'KITTI/vehicle_BEV_AP40_hard_strict': 0.0,

'KITTI/vehicle_2D_AP40_hard_strict': 0.0,

'KITTI/vehicle_3D_AP40_easy_loose': 0.0,

'KITTI/vehicle_BEV_AP40_easy_loose': 0.0,

'KITTI/vehicle_2D_AP40_easy_loose': 0.0,

'KITTI/vehicle_3D_AP40_moderate_loose': 0.0,

'KITTI/vehicle_BEV_AP40_moderate_loose': 0.0,

'KITTI/vehicle_2D_AP40_moderate_loose': 0.0,

'KITTI/vehicle_3D_AP40_hard_loose': 0.0,

'KITTI/vehicle_BEV_AP40_hard_loose': 0.0,

'KITTI/vehicle_2D_AP40_hard_loose': 0.0}

 

result 레포트가 이모양으로 나온건

Image 를 안넣어줬기 때문인 것으로 보입니다.

직접 result를 출력하는 코드를 짜줘야할 것 같습니다.