Update has_mask method for mmdet models

[Alias-z]

Solve the issue mentioned in #1053, which leads to empty instance segmentation results with mmdet models

[Alias-z]

Ummm, probably the CI test ran on an old style of mmdet config, leading to an error. I will try to make that compatible

[Alias-z]

@fcakyon

class MmdetTestConstants:
    MMDET_CASCADEMASKRCNN_MODEL_URL = "http://download.openmmlab.com/mmdetection/v2.0/cascade_rcnn/cascade_mask_rcnn_r50_fpn_1x_coco/cascade_mask_rcnn_r50_fpn_1x_coco_20200203-9d4dcb24.pth"
    MMDET_CASCADEMASKRCNN_MODEL_PATH = (
        "tests/data/models/mmdet/cascade_mask_rcnn/cascade_mask_rcnn_r50_fpn_1x_coco_20200203-9d4dcb24.pth"
    )
    MMDET_RETINANET_MODEL_URL = "http://download.openmmlab.com/mmdetection/v2.0/retinanet/retinanet_r50_fpn_2x_coco/retinanet_r50_fpn_2x_coco_20200131-fdb43119.pth"
    MMDET_RETINANET_MODEL_PATH = "tests/data/models/mmdet/retinanet/retinanet_r50_fpn_2x_coco_20200131-fdb43119.pth"
    MMDET_YOLOX_TINY_MODEL_URL = "https://download.openmmlab.com/mmdetection/v2.0/yolox/yolox_tiny_8x8_300e_coco/yolox_tiny_8x8_300e_coco_20211124_171234-b4047906.pth"
    MMDET_YOLOX_TINY_MODEL_PATH = "tests/data/models/mmdet/yolox/yolox_tiny_8x8_300e_coco_20211124_171234-b4047906.pth"

    MMDET_CASCADEMASKRCNN_CONFIG_PATH = "tests/data/models/mmdet/cascade_mask_rcnn/cascade-mask-rcnn_r50_fpn_1x_coco.py"
    MMDET_RETINANET_CONFIG_PATH = "tests/data/models/mmdet/retinanet/retinanet_r50_fpn_1x_coco.py"
    MMDET_YOLOX_TINY_CONFIG_PATH = "tests/data/models/mmdet/yolox/yolox_tiny_8xb8-300e_coco.py"

The test was indeed on mmdet v2.0, how do you like to change this?

[fcakyon]

@Alias-z you can drop the support for mmdet 2.0 and update it with mmdet 3.0 models/configs

[Alias-z]

Well, the CI test is tough ... @fcakyon

I initially used mmengine to combine all config inheritances into one config for the three test models; then I realized that there would many more code to modify. So i reverted back to current CI test configs, while updated has_mask method to be more universal:

Tested the new has_mask method on two object detectors and two instance segmentors:

1. yolox_tiny_8x8_300e_coco, returned False
2. 'retinanet_r50_fpn_1x_coco, returned False
3. cascade-mask-rcnn_r50_fpn_1x_coco, returned True
4. mask2former_swin-s-p4-w7-224_8xb2-lsj-50e_coco, returned True

So it should work now...

[fcakyon]

@Alias-z a very elegant solution 💯