PaddlePaddle / PaddleVideo

Awesome video understanding toolkits based on PaddlePaddle. It supports video data annotation tools, lightweight RGB and skeleton based action recognition model, practical applications for video tagging and sport action detection.
Apache License 2.0
1.52k stars 378 forks source link

针对attentionLSTM微调至少需要多少数据量? #238

Open Skye2099 opened 2 years ago

Skye2099 commented 2 years ago

想对自己的数据进行自定义打标预测,根据文档用tsn进行了特征提取,进行微调的话,我至少需要准备多少个带自定义标签的视频数据?

luyao-cv commented 7 months ago

具体到Attention LSTM模型微调,如果任务是高度特定且简单的,也许几十到几百个高质量标注的视频片段就足够了;