Support more datasets

[henrywang1]

I would like to make a proposal for the issues about supporting other datasets. e.g. #168 and #207.
#168 doesn't have a clear solution so far, and #207 doesn't support evaluating segmentation results.

Inspired by Detectron, I propose that adding a new config option, "DATASETS.FORCE_USE_JSON_ANNOTATION".
If the option is True, maskrcnn-benchmark will use COCO API to load the specified dataset and evaluate the result with the COCO styled AP. (i.e. AP, AP50, AP75, APs, APm, APl for bbox and segm)

In addition to the newly added option, I also

• Provide the steps to run on other datasets (maskrcnn_benchmark/data/Readme.md)
• Add the example configuration for Cityscapes
• Organize the codes for converting Cityscapes Datasets to COCO style
  • Reference: [1], [2]

You could try the below command options to evaluate the changes.

python train_net.py --config-file "configs/pascal_voc/e2e_mask_rcnn_R_50_FPN_1x.yaml" DATASETS.FORCE_USE_JSON_ANNOTATION True
python train_net.py --config-file "configs/e2e_faster_rcnn_R_50_FPN_1x.yaml" DATASETS.FORCE_USE_JSON_ANNOTATION True
python train_net.py --config-file "configs/e2e_mask_rcnn_R_50_FPN_1x.yaml" DATASETS.FORCE_USE_JSON_ANNOTATION True

Thanks !

[fmassa]

Thanks a lot for the PR, it will definitely be very helpful to the community!

I have left a few comments, let me know what you think.

The biggest comment is that I'd rather avoid having the new config flag, and instead I'd go with a new name for the dataset, like voc_2007_train_cocostyle.

Thoughts?

[henrywang1]

Hi @fmassa
I agree with all your comments, so I made some changes.
While I was modifying the code, I think that

Instead of having two entries in path_catalog.py

"voc_2012_train": ("voc/VOC2012", 'train'),
"voc_2012_train_cocostyle": ("voc/VOC2012/JPEGImages", "voc/VOC2012/Annotations/pascal_test2012.json"),

If we use dictionary to replace tuple, it will be more readable

"voc_2012_train": {
    "data_dir": "voc/VOC2012",
    "img_dir": "voc/VOC2012/JPEGImages",
    "ann_file": "voc/VOC2012/Annotations/pascal_train2012.json",
    "split": "train"
},

In addition, I think using code to append *_cocostyle entries to DATASETS is clearer.

coco_style_datasets = {
    key + "_cocostyle": value for key, value in DATASETS.items() if "coco" not in key
}	    }
DATASETS={**DATASETS, **coco_style_datasets}

What do you think about all the updates?
Btw, I will modify maskrcnn_benchmark/data/Readme.md, once we reach an agreement.

[fmassa]

This is looking great, thanks!

I have a few more comments, let me know what you think, I'm open to suggestions!

[henrywang1]

Thanks for the review and all the comments.
I've updated the code, the format and logic have changed to:

• voc_xxx and voc_xxx_cocostyle
• cityscapes_xxx_cocostyle

[fmassa]

Thanks a lot!