Different settings in config compared to maskrcnn_benchmark about detectron2 HOT 6 CLOSED

You need to look at the base configs for each config, e.g. https://github.com/facebookresearch/detectron2/blob/master/configs/Base-RCNN-FPN.yaml.

from detectron2.

Okay. I see that now. Thank you for your quick reply!

from detectron2.

Then, is the performance gap mainly coming from the multi-scale training? INPUT.MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)

from detectron2.

the performance gap

I don't know "what" performance gap you mean.

As the model zoo said, https://github.com/facebookresearch/detectron2/tree/master/configs/Detectron1-Comparisons has configs that are closer to Detectron1 settings.

from detectron2.

Sorry I mean the results in maskrcnn_benchmark repo: https://github.com/facebookresearch/maskrcnn-benchmark/blob/master/MODEL_ZOO.md

Now I see most of the config settings are the same except the INPUT.MIN_SIZE_TRAIN.

from detectron2.

INPUT.MIN_SIZE_TRAIN is also the same as maskrcnn-benchmark in https://github.com/facebookresearch/detectron2/tree/master/configs/Detectron1-Comparisons

As the above link says, the accuracy improvements come from a more correct implementation (such as a correct flip augmentation). The details will be shared later.

from detectron2.