Dear author, thanks for your great work. I have some questions about the instances and instance features.
- Is the maximum number of instances in the scene k1*k2 (number of nodes in coarse level times the number of nodes in fine level)? So, if I have a large scene with large number of instances in the scene, I need to set k1 or k2 larger. Have you experimented with datasets with large instance count in the scene? Will this lower the performance of the model?
- After the codebooks are trained, are the 6-d instance features not needed any more? For click-based selection, we can just read the instance index of the point and find the instance. For text queries, we can match the text's clip feature with clip features stored in the codebook for each instance.
Dear author, thanks for your great work. I have some questions about the instances and instance features.