Experiments on 3D Feature Visualization Comparison in the Paper

[Wting0106]

Thanks to the authors for the impressive work!
I'm wondering how the visualization results of the LangSplat 3D features presented in the paper were achieved. When I directly decompress the LangSplat 3D features into 512-dimensional vectors and match them with textual feature vectors, the results are quite poor, unlike the reasonably good results shown in the paper. Could the authors kindly clarify how this was implemented? Any advice would be greatly appreciated. Thanks!

[Sugar55888]

@Wting0106 Hi, I am facing the same problem. How do you restore the 3D features of each 3D Gaussian point of LangSplat to a 512-dimensional vector and match it with the text features?

[Wting0106]

@Wting0106 Hi, I am facing the same problem. How do you restore the 3D features of each 3D Gaussian point of LangSplat to a 512-dimensional vector and match it with the text features?

The 3D features obtained from LangSplat training were decoded to 512 dimensions using an auto-decoder and then matched with text. However, the results were very poor—far from what was shown in Figure 6 of the OpenGaussian paper. I'm not sure whether the problem lies in my matching method or if the 3D features trained by LangSplat are inherently of low quality.

[Sugar55888]

@Wting0106 I used the same approach as you, and the results were not good either. The indicators of miou are the same as those reported in the opengaussian paper. But I don't understand why the visualization effect of langsplat in Figure 6 is so good.

[donotlookatmyname]

@Sugar55888 @Wting0106 Hello, I've recently started exploring this field. I noticed that in the paper, the authors evaluated both Langsplat and LEGaussians using the same assessment methods across two datasets. In the comments section, I saw that you successfully reproduced the evaluation results for Langsplat using the authors' methods. Would it be possible to reach out to you to learn more about how you implemented this? Alternatively, I'd greatly appreciate any advice you could offer on the specific implementation details. Thank you very much!