训练时空动作检测 ，如何处理类别不均衡的问题

[yuweiApp]

What is the problem this feature will solve?

我使用slowfast_kinetics400-pretrained-r50_8xb8-8x8x1-20e_ava21-rgb.py配置文件，训练sit,stand,talk三个动作，其中stand标签比sit和talk多，训练的结果如下:
Overall Results:
Overall mAP: 0.3650
Person Movement mAP: 0.3650
Object Manipulation mAP: nan
Person Interaction mAP: nan
Class sit AP: 0.1524
Class stand AP: 0.8997
Class talk AP: 0.0429
03/09 09:59:35 - mmengine - INFO - Epoch(val) [20][49/49] mAP/overall: 0.3650 mAP/person_movement: 0.3650 mAP/object_manipulation: nan mAP/person_interaction: nan data_time: 0.0081 time: 0.1193

What is the feature?

我想大概是类别不均衡造成的原因，我该如何处理呢

What alternatives have you considered?

No response

[azhen01]

我也遇到了同样的问题，请问你现在有方法了吗？