S3DIS Dataset

[zzzyf01]

Hello, thank you for your amazing work! I am wondering if you could provide any code or guidance for processing S3DIS Dataset. Additionally, I noticed your model incorporates image captions—could you share the image captions of the S3DIS Dataset or details about the prompt you used to generate these captions? Thank you!

[Yanbo-23]

Thank you for your interest in our work! We generated the view-level captions for the dataset using the method described in PLA (Ding, 2022). Specifically, we utilized GPT-ViT2 to produce view-based captions such as "a kitchen scene with a window and a stove." Additionally, we have made the processed captions available for both the ScanNet and S3DIS datasets, which you can directly use here.
caption_view_s3dis_vit-gpt2-image-captioning_.json
caption_view_scannet_vit-gpt2-image-captioning_.json

Regarding the S3DIS dataset, consistent with prior works, we used Area 5 as the testing split, while the remaining scenes were used for training. The training process for S3DIS is the same as that for the ScanNet dataset, you can do that by modifying the scannet20/200 dataloader.