Inquiry Regarding Grasp Generation Discrepancy in Multi-View Reconstruction Point cloud

[Choi-jk]

Description:
Thanks for the excellent work on this project. I am deeply interested in its potential applications. Currently, I am integrating your model into my multi-view reconstruction method for grasp generation. However, I have observed an issue where the multi-view point cloud, accumulated from multiple depth images, produces a suboptimal number of grasps (zero or < 10) compared to the single-view point cloud generated from a single view depth image (~hundreds).

Question:
I am seeking clarification on why this discrepancy is occurring.

Grasp Generation Process for Multi-View Point Cloud:

1. Rendered depth map from multiple cameras
2. Converted each depth map to a point cloud (in world coordinates)
3. Combined all point clouds into one large point cloud
4. Created virtual cameras facing the center of the point cloud (assuming the object is at the center)
5. Converted the consolidated point cloud to each virtual camera's coordinates and used it as input for GraspNet

Grasp Generation Process for Single-View Point Cloud:

1. Rendered depth map from a single camera
2. Converted the depth map to a point cloud (in world coordinates)
3. Converted this point cloud to the camera's coordinates and used it as input for GraspNet

I would appreciate any insights or suggestions to address this issue. Your assistance is invaluable.
Thank you for your time and support.

[chenxi-wang]

Are the poses of your virtual cameras similar to the ones of the original data? The viewpoint may influence the performance.

[Choi-jk]

Yes, the virtual camera poses we employed closely resemble the ones I use for generating point clouds. Just to clarify, we also conducted tests using actual camera poses, and the results were similar.

[chenxi-wang]

I've not tested the baseline model on multi-view point clouds. Could the output candidates be filtered by collision detection?