预训练模型在不同线数激光雷达上(VLP16, VLP32)泛化性能不佳 (The pretrained model doesn't work well with 16- and 32-beam LiDAR data)

[yst1]

你好，我用你训练好的模型跑自己的16线雷达的数据，根据label输出对当前帧点云每个点区分静动态，最后输出静动态点云。但是效果是这样的，彩色的是动态点，请问可能是什么原因呢？

[Chen-Xieyuanli]

@yst1 你好，感谢使用我们的代码。从提供的实验结果来看，预训练的模型在不同激光雷达上的泛化能力不行。可行的解决方案是对预训练进行微调再训练（fine tune）。可以你的数据集中标注少量动静态物体标签，然后再训练一下，结果应该能够得到提升。

Q: This provided screenshot is the result of using the pre-trained model with the own collected 16-beam LiDAR data. As can be seen, the MOS results are not good. How to get better results on 16-beam LiDAR data.
A: The experiments show that the generalization of the pre-trained model is not very good for different types of LiDAR scanners. To make it work with 16-beam LiDAR data, one could label several binary labels on own collected data, and fine-tune the pre-trained model.

[yuhang9803]

你好，我用你训练好的模型跑自己的16线雷达的数据，根据label输出对当前帧点云每个点区分静动态，最后输出静动态点云。但是效果是这样的，彩色的是动态点，请问可能是什么原因呢？

你好，请问你是怎么把这个模型部署在ros上，然后输出ros格式的topic，然后在rviz上显示出来的呢？可以交流下吗？谢谢了

[Chrislzy1993]

@yuhang9803 ，修改输入输出接口，作者的输出好像是bin文件

[ypfsmile]

通过将32线数据，转换为需要的输入格式，基于预训练模型测试下来，发现几乎所有点都被认为是动态物体，有可能是模型泛化不好产生么，个人总感觉是数据输入错误。
By converting the vlp-32 data into the required input format and testing it based on the pre-training model, it is found that almost all points are considered dynamic objects. Is it possible that the model generalization is not good? Personally, I always feel that it is a data input error. Have you encountered similar problems?