How to improve the ability to understand videos with a small proportion of the target in the video

[ccqedq]

In this project, detection and video classification were not combined. If the person is far away from the camera and has a small proportion in the picture, is it necessary to concatenate detection and video classification? I understand that the detection model has stronger detection ability but slightly weaker classification ability. How can we improve the object detection model to enhance its classification ability