search

Jumpat / SegAnyGAussians

The official implementation of SAGA (Segment Any 3D GAussians)
Apache License 2.0
490 stars 36 forks source link

about downsample #87

Closed Zhao-Yian closed 4 days ago

Zhao-Yian commented 2 weeks ago

If I want to reproduce the metrics in the paper, what downsample should be set to? If it is set to 1, the memory usage is very high, but can the results be reproduced if it is set to 4/8?

Zhao-Yian commented 2 weeks ago

In addition, can you provide the results of each scene on the NVOS dataset? If convenient, can we discuss it in detail via email or WeChat?

Jumpat commented 2 weeks ago

If I want to reproduce the metrics in the paper, what downsample should be set to? If it is set to 1, the memory usage is very high, but can the results be reproduced if it is set to 4/8?

Hi, it's okay to downsample in practice since the 3D prior can help refine the segmentation results. Downsampling does not affect too much about the final performance. However, during training, an important issue is keeping the resolution of the rendered feature map and upsample the gt feature.

I can provide the per-scene results. You can leave the e-mail address or WeChat id and I will reach out.

Zhao-Yian commented 2 weeks ago

Thank you for your answer. Thank you for providing more results. My WeChat ID is Yian_Zhao