search

yangyangyang127 / Seg-NN

[CVPR2024 Hightlight] No Time to Train: Empowering Non-Parametric Networks for Few-shot 3D Scene Segmentation
78 stars 4 forks source link

About Cross-Correlation #7

Open leesb7426 opened 1 month ago

leesb7426 commented 1 month ago

I have some questions about the computing process of 'Cross-correlation' in L54-58 of 'models/quest.py'. In your paper, the cross-correlation is calculated by the inner product between the support feature (M, D) and the query feature (M, D). However, in your official code, the cross-correlation is calculated differently from what was mentioned in the paper. This is caused by a mix of information between instances in the batch due to the reshape in L54. If you have any intentions in that calculation process, can you tell me?

Furthermore, the results are significantly lower if the code is modified using the computational process mentioned in the paper.

yangyangyang127 commented 1 month ago

Thanks for your carefulness. You find a really important problem. This is a very weird problem. I also find this problem in PAPFZ-3D. Because it can improve the performance, I did not correct this code.

leesb7426 commented 1 month ago

If so, it means what is said in the paper and what is actually working differently, which seems like a serious problem. I think follow-up measures are necessary for the development of this field.

yangyangyang127 commented 1 month ago

Actually, this permutation trick is helpful to the performance, but you still need the methods described in the paper to achieve higher performance.

---- Replied Message ---- | From | @.> | | Date | 07/24/2024 22:26 | | To | @.> | | Cc | @.>@.> | | Subject | Re: [yangyangyang127/Seg-NN] About Cross-Correlation (Issue #7) |

If so, it means what is said in the paper and what is actually working differently, which seems like a serious problem. I think follow-up measures are necessary for the development of this field.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>