Can Spatial Temporal Action Detection SlowFast model predict bbox？

[ZJU-lishuang]

Can Spatial Temporal Action Detection SlowFast model predict bbox？
If it can predict bbox,how to do it?

[irvingzhang0512]

No. bboxes are the inputs of spatial temporal action detection model.

[irvingzhang0512]

Is there a better model than slowfast in Spatial Temporal Action Detection?
bboxes as the inputs of spatial temporal action detection is a sad thing.

In order to get better bboxes, we'd better use another model to detect humans.
For now, slowfast is the best in mmaction2

[ZJU-lishuang]

Is there a better model in Spatial Temporal Action Detection？
not only in mmaction2.
I didn't find one.

[ZJU-lishuang]

MOC-detector is not good in ava dataset.

[irvingzhang0512]

try alphaction, good-looking, hard to train and install.

[ZJU-lishuang]

Now I try to convert custom dataset to ava format.I got a csv like ava.
And I find the csv file already has the box coordinates.
What is the function of proposals in ava_dense_proposals_train.FAIR.recall_93.9.pkl during training slwofast?