What is the input image resolution of different models in the Detectron2 Model Zoo?

[melodjia]

What is the input image resolution (image size) of different models in the Detectron2 Model Zoo? For example, when using an image with a 4K resolution as input, to what size does it transform, for example, in the R50-FPN model or another model? Where can I find this information?

[sukjunhwang]

Unless changes from the default config file (detectron2/config/defaults.py) are noted, the default test pipeline resizes the input images with the parameters below.

_C.INPUT.MIN_SIZE_TEST = 800
_C.INPUT.MAX_SIZE_TEST = 1333

Please refer to the implementation of ResizeShortestEdge in detectron2/data/transforms/augmentation_impl.py to see how the images get transformed with the parameters.

[melodjia]

Если не будут отмечены изменения из файла конфигурации по умолчанию (detectron2/config/defaults.py), тестовый конвейер по умолчанию изменяет размер входных образов с параметрами ниже.

_C.INPUT.MIN_SIZE_TEST = 800
_C.INPUT.MAX_SIZE_TEST = 1333

Пожалуйста, ознакомьтесь с реализацией **ResizeShortestEdge **в detectron2/data/transforms/augmentation_impl.py, чтобы узнать, как изображения трансформируются с помощью параметров.

You did not understand me. Different pre-trained models have different input sizes because already have a architecture defined. Image resolution must match the input size of the pre-trained. My question is: What is the input image resolution of different models in the Detectron2 Model Zoo?

[ppwwyyxx]

Image resolution must match the input size of the pre-trained.

No. You can use any image resolution that you think is reasonable for any models in detectron2 model zoo.

The default resolution that we use is answered above by #3461 (comment)

[mahilaMoghadami]

Unless changes from the default config file (detectron2/config/defaults.py) are noted, the default test pipeline resizes the input images with the parameters below.

_C.INPUT.MIN_SIZE_TEST = 800
_C.INPUT.MAX_SIZE_TEST = 1333

Please refer to the implementation of ResizeShortestEdge in detectron2/data/transforms/augmentation_impl.py to see how the images get transformed with the parameters.

hello
you mean that for image with size (900*900), which resizing is done in detectrone2?