Dear Author(s)/Research Team,
First of all, congratulations to you and your team on your excellent work on SAMPart3D.
I am currently studying your research and am very interested in some of the technical details. I have a question regarding the large-scale pre-training stage and would be very grateful for your guidance.
In the section "Distilling 2D Visual Features to 3D Backbone," you mention sampling N points from each 3D mesh surface to extract their features. I have carefully reviewed your paper and the associated open-source code, but I was unable to find the specific value used for N.
Therefore, I would like to ask the following questions:
-
What was the specific value of the sample count N during the pre-training stage?
-
Does this value of N need to be consistent with the 15,000 points mentioned in the subsequent "Sample-specific Fine-tuning" stage?
-
If the number of points in these two stages is inconsistent (e.g., N in pre-training is not equal to 15k), would this lead to a significant drop in model performance?
-
Alternatively, is it possible that the value of N was not fixed during pre-training, or that there is no necessary correlation between the number of points sampled in pre-training versus fine-tuning?
Thank you for your time and for clarifying this detail. I look forward to your response.
Best regards.
Dear Author(s)/Research Team,
First of all, congratulations to you and your team on your excellent work on SAMPart3D.
I am currently studying your research and am very interested in some of the technical details. I have a question regarding the large-scale pre-training stage and would be very grateful for your guidance.
In the section "Distilling 2D Visual Features to 3D Backbone," you mention sampling N points from each 3D mesh surface to extract their features. I have carefully reviewed your paper and the associated open-source code, but I was unable to find the specific value used for N.
Therefore, I would like to ask the following questions:
What was the specific value of the sample count N during the pre-training stage?
Does this value of N need to be consistent with the 15,000 points mentioned in the subsequent "Sample-specific Fine-tuning" stage?
If the number of points in these two stages is inconsistent (e.g., N in pre-training is not equal to 15k), would this lead to a significant drop in model performance?
Alternatively, is it possible that the value of N was not fixed during pre-training, or that there is no necessary correlation between the number of points sampled in pre-training versus fine-tuning?
Thank you for your time and for clarifying this detail. I look forward to your response.
Best regards.