Lower mAP than reported when reproducing DSPDet3D on ScanNet-md40

[linQian99]

Hi, thank you for your excellent work on DSPDet3D!

I'm trying to reproduce your results on the ScanNet-md40 benchmark using the official configuration provided in your paper and repo. However, I noticed that my results are slightly lower than reported.
Here are my key settings:

---

voxel_size = 0.01

pts_prune_threshold = 100000

r = 7, volume_threshold = 27

prune_threshold = 0.3

Backbone: DSPBackbone(depth=34, max_channels=128, pool=False)

Head: DSPHead(in_channels=(64, 128, 128, 128), out_channels=128, n_classes=22, n_reg_outs=6)

Training: 12 epochs with RepeatDataset(times=10), batch size 4 per GPU

---

The training and test pipelines match those from the paper, and I'm using the ScanNet-md40 split with 22 classes. However, the performance (especially mAP@0.25) is around 1-2% lower than the result reported in Table 1 of the paper, even under the same pruning threshold setting (e.g., τ=0.3).

Could you kindly advise on:

Whether there are any additional tricks or tuning (e.g., post-processing, test-time settings) not mentioned in the paper?

Which seed(s) or randomness control were used to stabilize the reported results?

Was the 25-trial average used for table reporting, or best of several runs?

Any insights would be greatly appreciated!

Thanks again for your contribution!

voxel_size = .01
n_points = 100000

model = dict(
    type='DSPDet3D',
    voxel_size=voxel_size,
    backbone=dict(type='DSPBackbone', in_channels=3, max_channels=128, depth=34,  pool=False, norm='batch'),
    head=dict(
        type='DSPHead',
        in_channels=(64, 128, 128, 128),
        out_channels=128,
        n_reg_outs=6,
        n_classes=22,
        voxel_size=voxel_size,
        pts_prune_threshold=100000,
        assigner=dict(
            type='DSPAssigner',
            top_pts_threshold=6,
        ),
        assign_type='volume',
        volume_threshold=27,
        r=7,
        prune_threshold=0.3,
        bbox_loss=dict(type='AxisAlignedIoULoss2', mode='diou', reduction='none')),
    train_cfg=dict(),
    test_cfg=dict(nms_pre=1000, iou_thr=.5, score_thr=.01))

optimizer = dict(type='AdamW', lr=.001, weight_decay=.0001)
optimizer_config = dict(grad_clip=dict(max_norm=10, norm_type=2))
lr_config = dict(policy='step', warmup=None, step=[8, 11])
runner = dict(type='EpochBasedRunner', max_epochs=12)
custom_hooks = [dict(type='EmptyCacheHook', after_iter=True)]

checkpoint_config = dict(interval=1, max_keep_ckpts=12)
log_config = dict(
    interval=50,
    hooks=[
        dict(type='TextLoggerHook'),
        # dict(type='TensorboardLoggerHook')
])
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = None
load_from = None
resume_from = None
workflow = [('train', 1)]

n_points = 100000
dataset_type = 'ScanNetDataset'
data_root = 'data/ScanNet-md40/'
class_names = ('bathtub', 'bed', 'bench', 'bookshelf', 'bottle', 'chair', 'cup', 'curtain', 'desk', 'door', 'dresser',
                 'keyboard', 'lamp', 'laptop', 'monitor', 'night_stand', 'plant', 'sofa', 'stool', 'table', 'toilet',
                 'wardrobe')
train_pipeline = [
    dict(
        type='LoadPointsFromFile',
        coord_type='DEPTH',
        shift_height=False,
        use_color=False,
        load_dim=6,
        use_dim=[0, 1, 2]),
    dict(type='LoadAnnotations3D'),
    dict(type='GlobalAlignment', rotation_axis=2),
    # we do not sample 100k points for scannet, as very few scenes have
    # significantly more then 100k points. so we sample 33 to 100% of them
    dict(type='PointSample', num_points=n_points),
    dict(
        type='RandomFlip3D',
        sync_2d=False,
        flip_ratio_bev_horizontal=.5,
        flip_ratio_bev_vertical=.5),
    dict(
        type='GlobalRotScaleTrans',
        rot_range=[-.02, .02],
        scale_ratio_range=[.9, 1.1],
        translation_std=[.1, .1, .1],
        shift_height=False),
    # dict(type='NormalizePointsColor', color_mean=None),
    dict(type='DefaultFormatBundle3D', class_names=class_names),
    dict(type='Collect3D', keys=['points', 'gt_bboxes_3d', 'gt_labels_3d'])
]
test_pipeline = [
    dict(
        type='LoadPointsFromFile',
        coord_type='DEPTH',
        shift_height=False,
        use_color=False,
        load_dim=6,
        use_dim=[0, 1, 2]),
    dict(type='GlobalAlignment', rotation_axis=2),
    dict(
        type='MultiScaleFlipAug3D',
        img_scale=(1333, 800),
        pts_scale_ratio=1,
        flip=False,
        transforms=[
            # we do not sample 100k points for scannet, as very few scenes have
            # significantly more then 100k points. so it doesn't affect inference
            # time and we ca accept all points
            # dict(type='PointSample', num_points=n_points),
            # dict(type='NormalizePointsColor', color_mean=None),
            dict(
                type='DefaultFormatBundle3D',
                class_names=class_names,
                with_label=False),
            dict(type='Collect3D', keys=['points'])
        ])
]
data = dict(
    samples_per_gpu=4,
    workers_per_gpu=4,
    train=dict(
        type='RepeatDataset',
        times=10,
        dataset=dict(
            type=dataset_type,
            data_root=data_root,
            ann_file=data_root + 'scannet_infos_train.pkl',
            pipeline=train_pipeline,
            filter_empty_gt=False,
            classes=class_names,
            box_type_3d='Depth')),
    val=dict(
        type=dataset_type,
        data_root=data_root,
        ann_file=data_root + 'scannet_infos_val.pkl', 
        pipeline=test_pipeline,
        classes=class_names,
        test_mode=True,
        box_type_3d='Depth'),
    test=dict(
        type=dataset_type,
        data_root=data_root,
        ann_file=data_root + 'scannet_infos_val.pkl',
        pipeline=test_pipeline,
        classes=class_names,
        test_mode=True,
        box_type_3d='Depth'))