Mmdetection: About multiscale training

[iceload123]

I found that the repo has already supported using different image scales in the training.
This should be included in data augmentation, then what does multiscale training mean?
I think multiscale training should has similar format to input image pyramid, which has not been supported by this repo yet?

[hellock]

You only need to modify the data.train.img_scale field in the config file.
For example, you can use a list of scales [(1333, 800), (1666, 1000)] so that each image will randomly select a scale between (1333, 800) and (1666, 1000).

data = dict(
    imgs_per_gpu=2,
    workers_per_gpu=2,
    train=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/instances_train2017.json',
        img_prefix=data_root + 'train2017/',
        img_scale=[(1333, 800), (1666, 1000)],
        img_norm_cfg=img_norm_cfg,
        size_divisor=32,
        flip_ratio=0.5,
        with_mask=True,
        with_crowd=True,
        with_label=True),
    val=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/instances_val2017.json',
        img_prefix=data_root + 'val2017/',
        img_scale=(1333, 800),
        img_norm_cfg=img_norm_cfg,
        size_divisor=32,
        flip_ratio=0,
        with_mask=True,
        with_crowd=True,
        with_label=True),
    test=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/instances_val2017.json',
        img_prefix=data_root + 'val2017/',
        img_scale=(1333, 800),
        img_norm_cfg=img_norm_cfg,
        size_divisor=32,
        flip_ratio=0,
        with_mask=False,
        with_label=False,
        test_mode=True))

[iceload123]

You only need to modify the data.train.img_scale field in the config file.
For example, you can use a list of scales [(1333, 800), (1666, 1000)] so that each image will randomly select a scale between (1333, 800) and (1666, 1000).

data = dict(
    imgs_per_gpu=2,
    workers_per_gpu=2,
    train=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/instances_train2017.json',
        img_prefix=data_root + 'train2017/',
        img_scale=[(1333, 800), (1666, 1000)],
        img_norm_cfg=img_norm_cfg,
        size_divisor=32,
        flip_ratio=0.5,
        with_mask=True,
        with_crowd=True,
        with_label=True),
    val=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/instances_val2017.json',
        img_prefix=data_root + 'val2017/',
        img_scale=(1333, 800),
        img_norm_cfg=img_norm_cfg,
        size_divisor=32,
        flip_ratio=0,
        with_mask=True,
        with_crowd=True,
        with_label=True),
    test=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/instances_val2017.json',
        img_prefix=data_root + 'val2017/',
        img_scale=(1333, 800),
        img_norm_cfg=img_norm_cfg,
        size_divisor=32,
        flip_ratio=0,
        with_mask=False,
        with_label=False,
        test_mode=True))

Thanks, but my question is "what does multiscale training means".
I am not sure if simply selecting different resize scale in training phase can be called multiscale training. I notice that in your poster, almost 2% mAP is gained by multiscale training, but simply setting training scale to [(1667,1000),(1333,800),(1000,600)] brings no benefits according to my experiments.

[hellock]

The term of "multiscale training" is adopted in many papers, which indicates resizing images to different scales at each iteration. In the challenge, we use the setting [(400, 1600), (1400, 1600)] which means the short edge are randomly sampled from 400~1400, and the long edge is fixed as 1600.

[iceload123]

The term of "multiscale training" is adopted in many papers, which indicates resizing images to different scales at each iteration. In the challenge, we use the setting [(400, 1600), (1400, 1600)] which means the short edge are randomly sampled from 400~1400, and the long edge is fixed as 1600.

Thanks.