search

HXMap / MapQR

[ECCV 2024] This is the official implementation of MapQR, an end-to-end method with an emphasis on enhancing query capabilities for constructing online vectorized maps.
https://arxiv.org/abs/2402.17430
MIT License
144 stars 8 forks source link

reproduced results #11

Open missTL opened 1 month ago

missTL commented 1 month ago

Why do our reproduced results always stabilize around 65.0, compared to 66.4 in your paper. The configuration is as follows: `base = [ '../datasets/custom_nus-3d.py', '../base/default_runtime.py' ] # plugin = True plugin_dir = 'projects/mmdet3d_plugin/'

If point cloud range is changed, the models should also change their point

cloud range accordingly

point_cloud_range = [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0]

point_cloud_range = [-15.0, -30.0,-2.0, 15.0, 30.0, 2.0] voxel_size = [0.15, 0.15, 4.0] dbound=[1.0, 35.0, 0.5]

grid_config = { 'x': [-30.0, -30.0, 0.15], # useless 'y': [-15.0, -15.0, 0.15], # useless 'z': [-10, 10, 20], # useless 'depth': [1.0, 35.0, 0.5], # useful }

img_norm_cfg = dict( mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)

For nuScenes we usually do 10-class detection

class_names = [ 'car', 'truck', 'construction_vehicle', 'bus', 'trailer', 'barrier', 'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone' ]

map has classes: divider, ped_crossing, boundary

map_classes = ['divider', 'ped_crossing','boundary']

fixed_ptsnum_per_line = 20

map_classes = ['divider',]

num_vec=50 fixed_ptsnum_per_gt_line = 20 # now only support fixed_pts > 0 fixed_ptsnum_per_pred_line = 20 eval_use_same_gt_sample_num_flag=True num_map_classes = len(map_classes)

input_modality = dict( use_lidar=False, use_camera=True, use_radar=False, use_map=False, use_external=True)

dim = 256 _posdim = dim//2 _ffndim = dim*2 _numlevels = 1 _num_points_inpillar = 8

bevh = 50

bevw = 50

bevh = 200 bevw = 100 queue_length = 1 # each sequence contains queue_length frames.

aux_seg_cfg = dict( use_aux_seg=True, bev_seg=True, pv_seg=True, seg_classes=1, feat_down_sample=32, pv_thickness=1, )

model = dict( type='MapTRv2', use_grid_mask=True, video_test_mode=False, pretrained=dict(img='ckpts/resnet50-19c8e357.pth'), img_backbone=dict( type='ResNet', depth=50, num_stages=4, out_indices=(3,), frozen_stages=1, norm_cfg=dict(type='BN', requires_grad=False), norm_eval=True, style='pytorch'), img_neck=dict( type='FPN', in_channels=[2048], out_channels=dim, start_level=0, add_extra_convs='on_output', num_outs=_numlevels, relu_before_extra_convs=True), pts_bbox_head=dict( type='MapTRv2Head', bev_h=bevh, bev_w=bevw, num_query=900, num_vec_one2one=100, num_vec_one2many=600, k_one2many=6, num_pts_per_vec=fixed_ptsnum_per_pred_line, # one bbox num_pts_per_gt_vec=fixed_ptsnum_per_gt_line, dir_interval=1,

query_embed_type='instance_pts',

    query_embed_type='instance',
    transform_method='minmax',
    gt_shift_pts_pattern='v2',
    num_classes=num_map_classes,
    in_channels=_dim_,
    sync_cls_avg_factor=True,
    with_box_refine=True,
    as_two_stage=False,
    code_size=2,
    code_weights=[1.0, 1.0, 1.0, 1.0],
    aux_seg=aux_seg_cfg,
    # z_cfg=z_cfg,
    transformer=dict(
        type='MapTRPerceptionTransformer',
        rotate_prev_bev=True,
        use_shift=True,
        use_can_bus=True,
        embed_dims=_dim_,
        encoder=dict(
            type='BEVFormerEncoder',
            num_layers=3,
            pc_range=point_cloud_range,
            num_points_in_pillar=_num_points_in_pillar_,
            return_intermediate=False,
            with_height_refine=True,
            transformerlayers=dict(
                type='BEVFormerLayer',
                attn_cfgs=[
                    dict(
                        type='TemporalSelfAttention',
                        embed_dims=_dim_,
                        num_levels=1),
                    dict(
                       type='HeightKernelAttention',
                       pc_range=point_cloud_range,
                       num_points_in_pillar=_num_points_in_pillar_,
                       attention=dict(
                           type='MSDeformableAttentionKernel',
                           embed_dims=_dim_,
                           num_heads=_num_points_in_pillar_,
                           dilation=1,
                           kernel_size=(2, 4),
                           num_levels=_num_levels_),
                       embed_dims=_dim_,
                    )
                ],
                feedforward_channels=_ffn_dim_,
                ffn_dropout=0.1,
                operation_order=('self_attn', 'norm', 'cross_attn', 'norm',
                                 'ffn', 'norm'))),
        decoder=dict(
            type='MapTRDecoder',
            num_layers=6,
            return_intermediate=True,
            query_pos_embedding='instance',
            num_pts_per_vec=fixed_ptsnum_per_pred_line,
            transformerlayers=dict(
                type='DetrTransformerDecoderLayer',
                attn_cfgs=[
                    dict(
                        type='MultiheadAttention',
                        embed_dims=_dim_,
                        num_heads=8,
                        dropout=0.1),
                    dict(
                        type='InstancePointAttention',
                        embed_dims=_dim_,
                        num_levels=1,
                        num_pts_per_vec=fixed_ptsnum_per_pred_line,
                    ),
                ],
                feedforward_channels=_ffn_dim_,
                ffn_dropout=0.1,
                operation_order=('self_attn', 'norm', 'cross_attn', 'norm',
                                 'ffn', 'norm'))
        )),
    bbox_coder=dict(
        type='MapTRNMSFreeCoder',
        # post_center_range=[-61.2, -61.2, -10.0, 61.2, 61.2, 10.0],
        post_center_range=[-20, -35, -20, -35, 20, 35, 20, 35],
        pc_range=point_cloud_range,
        max_num=50,
        voxel_size=voxel_size,
        num_classes=num_map_classes),
    positional_encoding=dict(
        type='LearnedPositionalEncoding',
        num_feats=_pos_dim_,
        row_num_embed=bev_h_,
        col_num_embed=bev_w_,
        ),
    loss_cls=dict(
        type='FocalLoss',
        use_sigmoid=True,
        gamma=2.0,
        alpha=0.25,
        loss_weight=2.0),
    loss_bbox=dict(type='L1Loss', loss_weight=0.0),
    loss_iou=dict(type='GIoULoss', loss_weight=0.0),
    loss_pts=dict(type='PtsL1Loss', 
                  loss_weight=5.0),
    loss_dir=dict(type='PtsDirCosLoss', loss_weight=0.005),
    loss_seg=dict(type='SimpleLoss', 
        pos_weight=4.0,
        loss_weight=1.0),
    loss_pv_seg=dict(type='SimpleLoss', 
                pos_weight=1.0,
                loss_weight=2.0),),
# model training and testing settings
train_cfg=dict(pts=dict(
    grid_size=[512, 512, 1],
    voxel_size=voxel_size,
    point_cloud_range=point_cloud_range,
    out_size_factor=4,
    assigner=dict(
        type='MapTRAssigner',
        cls_cost=dict(type='FocalLossCost', weight=2.0),
        reg_cost=dict(type='BBoxL1Cost', weight=0.0, box_format='xywh'),
        # reg_cost=dict(type='BBox3DL1Cost', weight=0.25),
        # iou_cost=dict(type='IoUCost', weight=1.0), # Fake cost. This is just to make it compatible with DETR head.
        iou_cost=dict(type='IoUCost', iou_mode='giou', weight=0.0),
        pts_cost=dict(type='OrderedPtsL1Cost', 
                  weight=5),
        pc_range=point_cloud_range))))

dataset_type = 'CustomNuScenesOfflineLocalMapDataset' data_root = file_client_args = dict(backend='disk')

train_pipeline = [ dict(type='LoadMultiViewImageFromFiles', to_float32=True), dict(type='RandomScaleImageMultiViewImage', scales=[0.5]), dict(type='PhotoMetricDistortionMultiViewImage'), dict(type='NormalizeMultiviewImage', **img_norm_cfg), dict( type='LoadPointsFromFile', coord_type='LIDAR', load_dim=5, use_dim=5, file_client_args=file_client_args), dict(type='CustomPointToMultiViewDepth', downsample=1, grid_config=grid_config), dict(type='PadMultiViewImageDepth', size_divisor=32), dict(type='DefaultFormatBundle3D', with_gt=False, with_label=False,class_names=map_classes), dict(type='CustomCollect3D', keys=['img', 'gt_depth']) ]

test_pipeline = [ dict(type='LoadMultiViewImageFromFiles', to_float32=True), dict(type='RandomScaleImageMultiViewImage', scales=[0.5]), dict(type='NormalizeMultiviewImage', **img_norm_cfg),

dict(
    type='MultiScaleFlipAug3D',
    img_scale=(1600, 900),
    pts_scale_ratio=1,
    flip=False,
    transforms=[
        dict(type='PadMultiViewImage', size_divisor=32),
        dict(
            type='DefaultFormatBundle3D', 
            with_gt=False, 
            with_label=False,
            class_names=map_classes),
        dict(type='CustomCollect3D', keys=['img'])
    ])

]

data = dict( samples_per_gpu=4, workers_per_gpu=4, # TODO train=dict( type=dataset_type, data_root=data_root, ann_file=data_root + 'nuscenes_map_infos_temporal_train.pkl', pipeline=train_pipeline, classes=class_names, modality=input_modality, aux_seg=aux_seg_cfg, test_mode=False, use_valid_flag=True, bev_size=(bevh, bevw), pc_range=point_cloud_range, fixed_ptsnum_per_line=fixed_ptsnum_per_gt_line, eval_use_same_gt_sample_num_flag=eval_use_same_gt_sample_num_flag, padding_value=-10000, map_classes=map_classes, queue_length=queue_length,

we use box_type_3d='LiDAR' in kitti and nuscenes dataset

    # and box_type_3d='Depth' in sunrgbd and scannet dataset.
    box_type_3d='LiDAR'),
val=dict(
    type=dataset_type,
    data_root=data_root,
    ann_file=data_root + 'nuscenes_map_infos_temporal_val.pkl',
    map_ann_file=data_root + 'nuscenes_map_anns_val.json',
    pipeline=test_pipeline,  bev_size=(bev_h_, bev_w_),
    pc_range=point_cloud_range,
    fixed_ptsnum_per_line=fixed_ptsnum_per_gt_line,
    eval_use_same_gt_sample_num_flag=eval_use_same_gt_sample_num_flag,
    padding_value=-10000,
    map_classes=map_classes,
    classes=class_names, modality=input_modality, samples_per_gpu=1),
test=dict(
    type=dataset_type,
    data_root=data_root,
    ann_file=data_root + 'nuscenes_map_infos_temporal_val.pkl',
    map_ann_file=data_root + 'nuscenes_map_anns_val.json',
    pipeline=test_pipeline, 
    bev_size=(bev_h_, bev_w_),
    pc_range=point_cloud_range,
    fixed_ptsnum_per_line=fixed_ptsnum_per_gt_line,
    eval_use_same_gt_sample_num_flag=eval_use_same_gt_sample_num_flag,
    padding_value=-10000,
    map_classes=map_classes,
    classes=class_names, 
    modality=input_modality),
shuffler_sampler=dict(type='DistributedGroupSampler'),
nonshuffler_sampler=dict(type='DistributedSampler')

)

optimizer = dict( type='AdamW', lr=6e-4, paramwise_cfg=dict( custom_keys={ 'img_backbone': dict(lr_mult=0.1), }), weight_decay=0.01)

optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))

learning policy

lr_config = dict( policy='CosineAnnealing', warmup='linear', warmup_iters=500, warmup_ratio=1.0 / 3, min_lr_ratio=1e-3) total_epochs = 24 evaluation = dict(interval=2, pipeline=test_pipeline, metric='chamfer', save_best='NuscMap_chamfer/mAP', rule='greater')

total_epochs = 50

evaluation = dict(interval=1, pipeline=test_pipeline)

runner = dict(type='EpochBasedRunner', max_epochs=total_epochs)

log_config = dict( interval=50, hooks=[ dict(type='TextLoggerHook'), dict(type='TensorboardLoggerHook') ]) fp16 = dict(loss_scale=512.) checkpoint_config = dict(max_keep_ckpts=1, interval=2) find_unused_parameters=True`

missTL commented 1 month ago

Why do our reproduced results always stabilize around 65.0, compared to 66.4 in your paper.

fishmarch commented 1 month ago

There may be some random for the final results. According to our experience, it is normal to get results around 66.0. How many ‘grad_norm: nan’ in your log? Too many ‘grad_norm: nan’ may cause a lower result, which should not happen.