Config about dense_head(One_Stage) + RoI_head(Two_Stage)

[zhongqiu1245]

Hi dear authors:
Sorry to bother you.
If we adopt One-Stage's method of judging positive and negative samples and sampling methods and add Two-Stage's RoI_Head, will we get a more reasonable structure and more powerful Two-Stage effect? (In other words, it uses the positive and negative sample determination method and sampling method of One-Stage dense_head to replace the positive and negative sample determination method and sampling method of Two-Stage RPN)
I mean this:
dense_head(One_Stage) + RoI_head(Two_Stage)
After all, it can have both the characteristics of One-stage and Two-stage Or mAP advantages, such as using autoassign+cascade_rcnn, FCOS + cascade_rcnn(the original paper of FCOS proves the feasibility of FCOS as RPN), etc.
Can you kindly post a config file for this model?
Finally, please forgive me for my bad English.
Thank you!

[shinya7y]

Considering the CenterNet2 paper (https://arxiv.org/abs/2103.07461), this feature will be useful.
I sent a PR #5061 , where most of dense heads have methods for RPN (simple_test_rpn, aug_test_rpn).
It will be a step toward supporting the feature.

[zhongqiu1245]

Hi @shinya7y
Thank you for your PR, your job is really amazing.
Could you provide some configs to demonstrate how to use it? (such as autoassign +cascadercnn, fcos+cascadercnn, retinanet + maskrcnn？)
It would be great if it could be provided.
Thank you！

[shinya7y]

Since more code updates are needed to support the feature, just changing configs doesn't work.
If supported, configs will be like below.

gfl_faster_rcnn_r50_fpn_1x_coco.py (example for GFL + Faster R-CNN)

_base_ = [
    '../_base_/models/faster_rcnn_r50_fpn.py',  # base two-stage config
    '../_base_/datasets/coco_detection.py',
    '../_base_/schedules/schedule_1x.py',
    '../_base_/default_runtime.py'
]

model = dict(
    rpn_head=dict(  # copy from bbox_head of single-stage config
        _delete_=True,  # ignore settings of RPNHead in base config
        type='GFLHead',  # TODO update for RPN
        num_classes=1,  # foreground class only
        in_channels=256,
        stacked_convs=4,
        feat_channels=256,
        anchor_generator=dict(
            type='AnchorGenerator',
            ratios=[1.0],
            octave_base_scale=8,
            scales_per_octave=1,
            strides=[8, 16, 32, 64, 128]),
        loss_cls=dict(
            type='QualityFocalLoss',
            use_sigmoid=True,
            beta=2.0,
            loss_weight=1.0),
        loss_dfl=dict(type='DistributionFocalLoss', loss_weight=0.25),
        reg_max=16,
        loss_bbox=dict(type='GIoULoss', loss_weight=2.0)),
    train_cfg=dict(
        # GFL uses ATSSAssigner
        rpn=dict(assigner=dict(_delete_=True, type='ATSSAssigner', topk=9)),
        rpn_proposal=dict(score_thr=-1.0)),
    test_cfg=dict(rpn=dict(score_thr=-1.0)))

# disable NumClassCheckHook
custom_hooks = None
# TODO update
# https://github.com/open-mmlab/mmdetection/blob/808472f2574dcb52a4b5ec819a2c8dfab1af4356/mmdet/datasets/utils.py#L139-L140

For now, we need to implement a new head GFLRPNHead referring to GFLHead and RPNHead.

[zhongqiu1245]

Thank you！

[zhaoxin111]

Has anyone reproduced the strategy mentioned in CenterNet2? I used mmdet to implement retinanet-faster-rcnn, mAP is close to the original faster rcnn.

[LMerCy]

@shinya7y Actually my ap is so slow use such config, how about your results?(i map all the class to 0 for simplify)

[LMerCy]

@shinya7y Actually my ap is so slow use such config, how about your results?(i map all the class to 0 for simplify)

@zhaoxin111 Is your config same to this?[change rcnn pos thr to 0.6, rpn nms thr to 0.7]

[zhaoxin111]

@shinya7y Actually my ap is so slow use such config, how about your results?(i map all the class to 0 for simplify)

@zhaoxin111 Is your config same to this?[change rcnn pos thr to 0.6, rpn nms thr to 0.7]

The mAP of faster-rcnn-retinanet is close to the paper, but the centernet* in my experiment is lower 6 point than original code.

[LMerCy]

@shinya7y Actually my ap is so slow use such config, how about your results?(i map all the class to 0 for simplify)

@zhaoxin111 Is your config same to this?[change rcnn pos thr to 0.6, rpn nms thr to 0.7]

The mAP of faster-rcnn-retinanet is close to the paper, but the centernet* in my experiment is lower 6 point than original code.

Did you use "RandomSampler" for retinanet train_cfg?

[zhaoxin111]

@shinya7y Actually my ap is so slow use such config, how about your results?(i map all the class to 0 for simplify)

@zhaoxin111 Is your config same to this?[change rcnn pos thr to 0.6, rpn nms thr to 0.7]

The mAP of faster-rcnn-retinanet is close to the paper, but the centernet* in my experiment is lower 6 point than original code.

Did you use "RandomSampler" for retinanet train_cfg?

No,because RetinaNet do not need sampler

[LMerCy]

@zhaoxin111 Thanks! i have reproduced the results, i made some mistake in my origin config.