question about making datasets for semantic-nerfw

[cezarbbb]

I recently find it's hard to get datasets properly for semantic-nerfw through reading https://github.com/ethanweber/sitcoms3D. I tried to make custom datasets by using ns-process-data, however, I find the generated image_4 and image_8 lost their original names(turned into frame**) resulting in error of ns-train(no such directory: image_4). Also, I don't know how to generate semantic_4 and semantic_8. Hope to see official documents about making semantic-nerfw datasets!

[ethanweber]

Hi @cezarbbb, you can look at my comment here in Issue #1275. We don't plan to prioritize semantic documentation/support at the moment, but we are very open to contributions! The idea would to modify the nerfstudio data parser to have semantic support (much like we do for the masks).

The semantic masks are produced with custom detectron2 code, but you could use any semantic segmentation model to create segmentation images.