How to build test-pipeline when RandomCrop is applied in train-pipeline

[ccomkhj]

Hi,

If the config below is the configuration in train/test_pipeline, model will see cropped image in training, but will see the whole image.
So, there's huge gap in understanding the scene (if crop_size and img_scale are significantly different.)

So, when I train model with SOLOV2 and HTC, performance is miserable. (less than 0.1 recall)
However, Mask-RCNN give decent output (more than 0.6 recall).

I think that Mask-RCNN give good result because the original paper included "RandomCrop".

What makes this difference?

[In mmseg, there's test_cfg = dict (mode='slide', crop_size= ... ), can this method applied to detection (instance segmentation) model?]

Here's te example of pipeline.

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
    dict(type='Resize', img_scale=img_scale, keep_ratio=True),
    dict(type='RandomFlip', flip_ratio=0.5),
    dict(
        type='Normalize',
        **img_norm_cfg),
    dict(
        type='RandomCrop',
        crop_size=crop_size,
        crop_type='absolute',
        recompute_bbox=True,
        allow_negative_crop=True),
    dict(
        type='FilterAnnotations', min_gt_bbox_wh=(1e-05, 1e-05), by_mask=True),
    dict(type='Pad', size_divisor=32),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks'])
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=img_scale,
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip', flip_ratio=0.5),
            dict(
                type='Normalize',
                **img_norm_cfg),
            dict(type='Pad', size_divisor=32),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img'])
        ])
]