Mmdetection: Finetuning on small dimension images

[tiberium24]

Hello,

I am wondering which parameters would you think is best to change in one of the config files when inputing low res images (due to their small size, 480*360). Definitely the resize option needs to be changed but what else would be needed to improve accuracy?

Thank you

[ZwwWayne]

You may further try multi-scale training to improve accuracy. A template is here.

[tiberium24]

Thanks. I observe however that you have this:

dict(
        type='Resize',
        img_scale=[(1600, 400), (1600, 1400)],
        multiscale_mode='range',
        keep_ratio=True)

Why is (1600,400) there two times?

Also, later you have:

test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(1333, 800),
        flip=False

Why is it in testing (1333,800)?

[ZwwWayne]

1. They are not the same. [(1600, 400), (1600, 1400)] means the short side ranges from 400 to 1400.
2. This is to make sure that the images have the same sizes during testing to avoid randomness.
   For the COCO dataset, a common practice is to use size (1333, 800). You can set a value suitable for your own dataset.

[tiberium24]

Thanks for your reply.

1. My mistake, I missed the second element.
2. I had the impression that the size to which you train needs to be the same when testing.

Do you have a similar example only for bounding boxes detector (without mask)?

[ZwwWayne]

The multi-scale setting works for both only detection or instance segmentation.