Inquiry about SUNRGBD Dataset Processing and High-Quality 3D Bbox Selection for Training

[averyjuuu]

Thank you for your excellent work, I am a researcher and currently working on a project involving the SUNRGBD dataset. I have been studying the codebase provided in your GitHub repository and found it to be immensely helpful.

I have a few questions related to the dataset processing and model training that I was hoping you could kindly address:

(1)SUNRGBD Dataset Processing: In the code, I noticed that there might be some additional preprocessing steps involved in handling the SUNRGBD dataset. Could you please provide some insights into the specific preprocessing steps applied to the dataset before training?

(2)High-Quality 3D Bbox Selection: I am particularly interested in understanding how you selected the high-quality 3D bounding boxes to be used during training. Are there any specific criteria or strategies you employed to ensure the selection of reliable 3D bounding boxes?

(3)Detic Segmentation Results Reliability: While exploring the segmentation results generated by Detic, I noticed promising performance. However, I would like to inquire about the reliability of these segmentation results in real-world scenarios. Could you share your observations or any evaluations you conducted regarding the accuracy and robustness of Detic-generated segmentations?

Your insights on these matters would significantly contribute to my research and understanding of the SUNRGBD dataset and your model implementation. I truly appreciate your time and assistance!