About the spatial-temporal patterns about tfusion HOT 3 CLOSED

Spatial temporal model couldn't extract cross dataset spatial temporal features, because st features across different camera topologies are different, which is hard to use for transfer across datasets.

We use vision model trained on source dataset, and use it to extract spartial temporal feature on target dataset, without any spatial temporal infomation from source dataset.

The key to sovle this cross dataset problems is sharing of the vision model, rather than the spatial-temporal model.

from tfusion.

ok, I understand, the spatial-temporal information is used to improve the source dataset performance, right?

from tfusion.

No, we only care about the performance on target dataset and we don't compute spatial-temporal pattern about source dataset. The spatial-temporal information can significantly improve the performance on target dataset compared to directly transfer using vision model. That's one point, and this will provide retraining data with better quality.

from tfusion.