Configuration for the

[xavibou]

Hello, I was able to train and test the H2RBox-v2 model using the provided config h2rbox_v2p_r50_fpn_1x_dota_le90.py. However, this accounts for the FCOS-based model with about 71.5 mAP in the DOTAv1 test set. I was wondering which config should be used in order to use Multi-scale (MS) and random rotation (RR) and achieve the higher results reported in the Table 2 of the article.

I am providing the exact config file i am using to achieve a performance of about 71.5 mAP:

dataset_type = 'DOTADataset'
data_root = 'data/DOTA/split_ss_dota1_0/'
img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(type='RResize', img_scale=(1024, 1024)),
    dict(
        type='RRandomFlip',
        flip_ratio=[0.25, 0.25, 0.25],
        direction=['horizontal', 'vertical', 'diagonal'],
        version='le90'),
    dict(
        type='Normalize',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375],
        to_rgb=True),
    dict(type='Pad', size_divisor=32),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(1024, 1024),
        flip=False,
        transforms=[
            dict(type='RResize'),
            dict(
                type='Normalize',
                mean=[123.675, 116.28, 103.53],
                std=[58.395, 57.12, 57.375],
                to_rgb=True),
            dict(type='Pad', size_divisor=64),
            dict(type='DefaultFormatBundle'),
            dict(type='Collect', keys=['img'])
        ])
]
data = dict(
    samples_per_gpu=2,
    workers_per_gpu=16,
    train=dict(
        type='DOTAWSOODDataset',
        ann_file='data/DOTA/split_ss_dota1_0/trainval/annfiles/',
        img_prefix='data/DOTA/split_ss_dota1_0/trainval/images/',
        pipeline=[
            dict(type='LoadImageFromFile'),
            dict(type='LoadAnnotations', with_bbox=True),
            dict(type='RResize', img_scale=(1024, 1024)),
            dict(
                type='RRandomFlip',
                flip_ratio=[0.25, 0.25, 0.25],
                direction=['horizontal', 'vertical', 'diagonal'],
                version='le90'),
            dict(
                type='Normalize',
                mean=[123.675, 116.28, 103.53],
                std=[58.395, 57.12, 57.375],
                to_rgb=True),
            dict(type='Pad', size_divisor=32),
            dict(type='DefaultFormatBundle'),
            dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
        ],
        version='le90'),
    val=dict(
        type='DOTAWSOODDataset',
        ann_file='data/DOTA/split_ss_dota1_0/trainval/annfiles/',
        img_prefix='data/DOTA/split_ss_dota1_0/trainval/images/',
        pipeline=[
            dict(type='LoadImageFromFile'),
            dict(
                type='MultiScaleFlipAug',
                img_scale=(1024, 1024),
                flip=False,
                transforms=[
                    dict(type='RResize'),
                    dict(
                        type='Normalize',
                        mean=[123.675, 116.28, 103.53],
                        std=[58.395, 57.12, 57.375],
                        to_rgb=True),
                    dict(type='Pad', size_divisor=64),
                    dict(type='DefaultFormatBundle'),
                    dict(type='Collect', keys=['img'])
                ])
        ],
        version='le90'),
    test=dict(
        type='DOTAWSOODDataset',
        ann_file='data/DOTA/split_ss_dota1_0/test/images/',
        img_prefix='data/DOTA/split_ss_dota1_0/test/images/',
        pipeline=[
            dict(type='LoadImageFromFile'),
            dict(
                type='MultiScaleFlipAug',
                img_scale=(1024, 1024),
                flip=False,
                transforms=[
                    dict(type='RResize'),
                    dict(
                        type='Normalize',
                        mean=[123.675, 116.28, 103.53],
                        std=[58.395, 57.12, 57.375],
                        to_rgb=True),
                    dict(type='Pad', size_divisor=64),
                    dict(type='DefaultFormatBundle'),
                    dict(type='Collect', keys=['img'])
                ])
        ],
        version='le90'))
evaluation = dict(interval=1, metric='mAP')
optimizer = dict(type='AdamW', lr=5e-05, betas=(0.9, 0.999), weight_decay=0.05)
optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
lr_config = dict(
    policy='step',
    warmup='linear',
    warmup_iters=500,
    warmup_ratio=0.3333333333333333,
    step=[8, 11])
runner = dict(type='EpochBasedRunner', max_epochs=12)
checkpoint_config = dict(interval=1)
log_config = dict(interval=50, hooks=[dict(type='TextLoggerHook')])
dist_params = dict(backend='nccl')
log_level = 'INFO'
load_from = None
resume_from = None
workflow = [('train', 1)]
opencv_num_threads = 0
mp_start_method = 'fork'
angle_version = 'le90'
model = dict(
    type='H2RBoxV2PDetector',
    backbone=dict(
        type='ResNet',
        depth=50,
        num_stages=4,
        out_indices=(0, 1, 2, 3),
        frozen_stages=1,
        zero_init_residual=False,
        norm_cfg=dict(type='BN', requires_grad=True),
        norm_eval=True,
        style='pytorch',
        init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50')),
    neck=dict(
        type='FPN',
        in_channels=[256, 512, 1024, 2048],
        out_channels=256,
        start_level=1,
        add_extra_convs='on_output',
        num_outs=5,
        relu_before_extra_convs=True),
    bbox_head=dict(
        type='H2RBoxV2PHead',
        num_classes=15,
        in_channels=256,
        stacked_convs=4,
        feat_channels=256,
        strides=[8, 16, 32, 64, 128],
        center_sampling=True,
        center_sample_radius=1.5,
        norm_on_bbox=True,
        centerness_on_reg=True,
        square_cls=[1, 9, 11],
        resize_cls=[1],
        scale_angle=False,
        bbox_coder=dict(type='DistanceAnglePointCoder', angle_version='le90'),
        loss_cls=dict(
            type='FocalLoss',
            use_sigmoid=True,
            gamma=2.0,
            alpha=0.25,
            loss_weight=1.0),
        loss_bbox=dict(type='IoULoss', loss_weight=1.0),
        loss_centerness=dict(
            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
        loss_ss_symmetry=dict(type='SmoothL1Loss', loss_weight=0.2, beta=0.1)),
    train_cfg=None,
    test_cfg=dict(
        nms_pre=2000,
        min_bbox_size=0,
        score_thr=0.05,
        nms=dict(iou_thr=0.01),
        max_per_img=2000))
work_dir = './work_dirs/h2rbox_v2p_r50_fpn_1x_dota_le90'
auto_resume = False
gpu_ids = range(0, 1)

Tanks in advance!

[yangxue0827]

ms means offline multi-scale image cropping, so you need to get a multiple scale dataset by running the following command before training.

python tools/data/dota/split/img_split.py --base-json \
  tools/data/dota/split/split_configs/ms_trainval.json

python tools/data/dota/split/img_split.py --base-json \
  tools/data/dota/split/split_configs/ms_test.json

[yangxue0827]

As for RR, please refer to https://github.com/yangxue0827/STAR-MMRotate/blob/48167d6067cacf4b2009b4c080112267e595b152/configs/roi_trans/roi_trans_r50_fpn_1x_dota_ms_rr_le90.py#L16-L22

[yangxue0827]

Actually, the original code of h2rbox-v2 is based on mmrotate1.x, while this code is based on mmrotate0.3.4. There are big differences between the two versions, so I cannot guarantee that you can reproduce it perfectly. It is recommended to use the original implementation code.