different result on suma++ and rangenet++ about semantic_suma

[Chen-Xieyuanli]

Hey @xdtzzz, thanks for using our code.

The noisy map results may be caused by the wrong prediction and inaccurate poes estimation.

For example, the points of a car were predicted to be both car (blue) and house (yellow) in the segmentation results.
When aggregating all the predictions together, it will become more obvious.

[xdtzzz]

谢谢您的回复！ 您的意思是，每个类别都必须要分割的十分精确才可以使suma+达到良好的效果吗？ 在rangenet运行，如果一个物体里大部分点云被成功的预测了，少量点云是错误的，就会导致suma+里面很多点云都是错误的情况吗 发自我的iPhone

…

------------------ Original ------------------ From: Xieyuanli Chen ***@***.***> Date: Thu,Dec 30,2021 3:54 PM To: PRBonn/semantic_suma ***@***.***> Cc: xdtzzz ***@***.***>, Mention ***@***.***> Subject: Re: [PRBonn/semantic_suma] different result on suma++ and rangenet++(Issue #52)

[Chen-Xieyuanli]

Q: Do you mean that SuMa++ needs very good semantic segmentation results to work well?
A: Yes, SuMa++ itself can not improve the semantic segmentation results, but only use the semantics to improve the SLAM results. Wrong predictions will lead to a wrong semantic map. We originally use RangeNet++ which achieved an IOU of 52.2 and worked well. The segmentation results seem not as good as the original RangeNet++. What's your segmentation performance in terms of IOU? The current SOTA one is of IOU 70, and you may use the SOTA semantics to get a better semantic map.

是的，SuMa++对semantic segmentation本身是没有提升的，只是利用语义信息来提高SLAM的精度，如果语义信息不准，语义地图也是会不准的。
原文用的rangenet++的IOU是52.2是可以工作的，当前的语义分割结果似乎比原文差，请问现在语义分割的IOU精度是多少？ 目前性能最好的语义分割算法性能是70IOU，可以尝试使用更好的语义分割算法。

[xdtzzz]

我现在语义分割的iou精度挺高的，因为标注的类别比较少，平均达到了0.8以上

[xdtzzz]

Results on rangenet++, most of them are correct. But on suma++, there are many wrong semantic points. For example, the ground points build on the top of the wall on suma++, but on rangenet++, the wall doesn't have ground points.
rangenet++:

suma++:

[xdtzzz]

where is "reopen" button.... # @Chen-Xieyuanli