How to resize gt-bbox when using RandomAffine for evaluating COCO val-set

Hi, @RangiLyu I meet the following question:

[Background] I am reproducting NanoDet-m-0.5x based on mmdet2. In official NanoDet code, you get resize_warpAffine_matrix first to resize the image when evaluating COCO val-set. To align the inference results with provided checkpoint weights, I must done the same thing as you do.

[My solution] I using the following config to try to reproduct:

test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(320, 320),
        flip=False,
        transforms=[
            # dict(type='Resize', keep_ratio=False),
            dict(type='RandomAffine', max_rotate_degree=0, max_translate_ratio=0, scaling_ratio_range=(1, 1), max_shear_degree=0, \
                border_val=(0, 0, 0),min_bbox_size=0,min_area_ratio=0,max_aspect_ratio=20000),
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            # dict(type='Pad', size_divisor=32),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]

However, I found that alough I have get the image I wanted, the corresponding gt bboxs are not resized. If I don’t use your checkpoint, and I train from scratch. the final COCO mAP is 13.0. In RandomAffine when test period, the bbox_fields is not exist, so what should I do.

[RangiLyu]

As I know, the resize method will not influence the performance so much.

[Senwang98]

@RangiLyu Sorry for disturb, my param and gflops are the same as your nanodet, but the final mAP is 13.0 without data aug when training. If I use RandomAffine and PhotoMetricDistortion when training, the final mAP is lower than 12.5. I have no idea about how to reproduct your nanodet based on mmdet2. If I want to resize test image just like your nanodet do (get resize affine matrix), what should I do?