On the source video with a relatively white skin color, the inference result will have a clear dark rectangular mask.
After trying it out, I found that as long as other videos of the characters in the source video (with similar backgrounds and lighting styles to the source video) are added to the training set, the model's performance will be much better.
I understand that incremental (continuous learning) data can be used to solve this problem. Does anyone know how to do it? Or is there any relevant code or reference materials available?
Thank you.
On the source video with a relatively white skin color, the inference result will have a clear dark rectangular mask. After trying it out, I found that as long as other videos of the characters in the source video (with similar backgrounds and lighting styles to the source video) are added to the training set, the model's performance will be much better. I understand that incremental (continuous learning) data can be used to solve this problem. Does anyone know how to do it? Or is there any relevant code or reference materials available? Thank you.
在肤色比较白的源视频上，推理结果会有一个明显的暗色矩形mask。 经过尝试我发现，只要将源视频里的人物的其他视频(背景和灯光等风格与源视频相似)添加到训练集，模型表现就会好很多。 我了解到可以对数据进行增量（持续学习），解决这个问题，请问有人知道该如何做吗？或者是否有相关的代码或参考资料？ 谢谢。