search

JacobYuan7 / DIN-Group-Activity-Recognition-Benchmark

[ICCV 2021] A new codebase containing various methods for Group Activity Recognition. Paper title: Spatio-Temporal Dynamic Inference Network for Group Activity Recognition.
MIT License
52 stars 15 forks source link

Generating annotations from saved model #13

Closed ch-its closed 2 years ago

ch-its commented 2 years ago

How can I generate annotations for input videos using a saved model without losing out on accuracy ??

JacobYuan7 commented 2 years ago

How can I generate annotations for input videos using a saved model without losing out on accuracy ??

Hi, thanks for your interest in our work, but I do not understand your question. Would you mind rephrasing it to make it clearer?

ch-its commented 2 years ago

Thanks for replying. I was asking how I could generate activity labels for a given video sequence using a saved model. It's sorted out now. On the other hand, the problem I'm facing now is that when I run the DIN model on Collective Activity Dataset, the accuracy actually decreases. How can I sort that out ??

JacobYuan7 commented 2 years ago

Thanks for replying. I was asking how I could generate activity labels for a given video sequence using a saved model. It's sorted out now. On the other hand, the problem I'm facing now is that when I run the DIN model on Collective Activity Dataset, the accuracy actually decreases. How can I sort that out ??

CAD is a dataset that easily overfits and encounters performance fluctuation because of the small dataset size. Usually, you need to use early stopping to achieve high performance.