[Feature] Support ActivityNetMeanAR

[hukkai]

WIP

[CLAassistant]

All committers have signed the CLA.

[C1rN09]

Is this PR ready for review, or still WIP? ^_^

[hukkai]

@C1rN09 Hi, it is ready

[C1rN09]

There should be 2 additional modifications:

1. Add unit tests for this metric. NOTE that 2 basic use cases should be covered: (1) stateless call, i.e. directly use metric's __call__ method to obtain result. (2) statefull call, i.e. call metric.add multiple (>=2) times and then metric.compute
2. Add api docs in metrics.rst

[hukkai]

@C1rN09 Not sure why the link failed, could you have a look?

[C1rN09]

@C1rN09 Not sure why the link failed, could you have a look?

This has been fixed in #86
You can rebase to newest master branch

[C1rN09]

LGTM. There should be a PR in mmaction2 to check the accuracy of this implementation.

[C1rN09]

Since PR #102, logger has became an argument in BaseMetric, so other metrics no longer need to define logger themselves.

[C1rN09]

Suggested change

	t_iou (np.ndarray): 1-dim array [n] /
	2-dim array [n x m] with IoU ratio.
	t_iou (np.ndarray): 1-dim array [n] /
	2-dim array [n x m] with IoU ratio.

[C1rN09]

Suggested change

	if target_segments.ndim != 2:
	raise ValueError('Dimension of target segments is incorrect')
	if candidate_segments_ndim not in [1, 2]:
	raise ValueError('Dimension of candidate segments is incorrect')
	if target_segments.ndim != 2:
	raise ValueError(f'Dimension of target segments is incorrect. Expected 2, got {target_segments.ndim}')
	if candidate_segments_ndim not in [1, 2]:
	raise ValueError(f'Dimension of candidate segments is incorrect. Expected 1 or 2, got {target_segments.ndim}')

[C1rN09]

When will these 2 if-statements be True? There has been

this_video_proposals = proposals_video_id[:, :2]

and

this_video_ground_truth = ground_truth_video_id[:, :2].astype(np.float32)

before.