[Enhancement] Support dynamic threshold range in eval_hmean

[gaotongxiao]

Motivation

The current implementation of eval_hmean only provides a parameter to set the minimum score with which the boundaries are used for evaluation. However, it always evaluates the filtered boundaries on a fixed set of threshold values [0.3, 0,4, ..., 0.9] no matter what the minimum score is, which is an inflexible and illogical design. There is also no entry in the config to customize such a behavior.

Modification

This PR replaces score_thr with min_score_thr, max_score_thr, step so that users can configure the search space via input parameters. IcdarDataset.evaluate() is modified to allow these three parameters to be customized through the config file.

For example, by adapting the following snippet in the config, one can evaluate the model's output on a list of boundary score thresholds [0.1, 0.2, 0.3, 0.4, 0.5] and get the best score from them.

evaluation = dict(
    interval=100,
    metric='hmean-iou',
    min_score_thr=0.1,
    max_score_thr=0.5,
    step=0.1)

BC-breaking (Optional)

None

[Mountchicken]

• eval_hmean is also used in TextDetDataset. We'd better modify it too.
• BTW, One thing that confuses me is that I didn't see these score_thr parameters that can be passed to eval_hmean in the config files, do users have to modify the source code to achieve this?

[gaotongxiao]

1. Sounds good
2. In fact, all the extra keys in the evaluation field of config are passed to dataset.evaluate. We just need to make sure dataset.evaluate parses the parameter correctly. See

   mmocr/mmocr/apis/train.py

   Line 146 in b4a9a87

   eval_cfg = cfg.get('evaluation', {})

   and https://github.com/open-mmlab/mmcv/blob/3a3514a54d59655e4a9ef88b72bb9757783da9db/mmcv/runner/hooks/evaluation.py#L363-L364