Realization of SF (Slicing Aided Fine-tuning) module and design of related parameters

[YipChakwai]

Hello sir, I would like to know whether the SF(Slicing Aided Fine-tuning) module is implemented by the train_pipeline in the mmdet_configs file?
train_pipeline = [ dict(type="LoadImageFromFile"), dict(type="LoadAnnotations", with_bbox=True), dict( type="AutoAugment", policies=[ [ dict(type="RandomCrop", crop_type="absolute_range", crop_size=(480, 960), allow_negative_crop=True), dict(type="Resize", img_scale=(1333, 800), keep_ratio=True), ], [ dict(type="RandomCrop", crop_type="absolute_range", crop_size=(480, 960), allow_negative_crop=True), dict(type="Resize", img_scale=(1333, 800), keep_ratio=True), ], [ dict(type="Resize", img_scale=(1333, 800), keep_ratio=True), ], ], ), dict(type="RandomFlip", flip_ratio=0.5), dict(type="Normalize", **img_norm_cfg), dict(type="Pad", size_divisor=32), dict(type="DefaultFormatBundle"), dict(type="Collect", keys=["img", "gt_bboxes", "gt_labels"]), ]
If yes, how should the parameter crop_size be designed? Is it designed according to the size of visdrone and xview dataset pictures?

[fcakyon]

Hello sir, I would like to know whether the SF(Slicing Aided Fine-tuning) module is implemented by the train_pipeline in the mmdet_configs file?

Yes.

If yes, how should the parameter crop_size be designed? Is it designed according to the size of visdrone and xview dataset pictures?

Yes.

You have to arrange the slice sizes considering your object to image ratio (if average size of objects/average size of images is very low you can use smaller slices.)

[YipChakwai]

Thank you for your answer. Should the size of the sliced pictures in SF and SAHI be the same? If the picture is randomly sliced during training (type="RandomCrop", crop_type="absolute_range"), how to design a fixed slice size when performing sahi predict?

[fcakyon]

You can use average size or max size during inference.

[YipChakwai]

Ok, I'll try. Thank you for your patience！