2d->3d problem

[wyy2546676927]

Thank you for your contribution. I currently have a confusion regarding your paper. You mentioned 3D-aware voxel-based projection for 2D images in the paper. Since I noticed that the provided code does not utilize depth information (depth is not used), I would like to ask: how do you manage to project a 2D image into 3D space without using the depth map? To further support my confusion, I noticed a specific detail in the real-world Go2 robot client code that seems to contradict the need for depth: The ROS topic subscription for the depth sensor (depth_sub) is commented out in the actual client implementation.Thank you for your answer.

self.depth_sub = self.create_subscription(Image, "/camera/camera/aligned_depth_to_color/image_raw",self.depth_callback, 1)

[yuqiang-yang]

We don't use the depth image in real-world experiments. We only do this voxel pooling in habitat simulation and verify its effectiveness.

[MuXiangWen]

We don't use the depth image in real-world experiments. We only do this voxel pooling in habitat simulation and verify its effectiveness.

Hello, thank you for releasing this project.

We have a question about the practical deployment of the system. In the demo, the robot is able to navigate toward a target object and stop in front of it. However, we are not sure how this behavior is achieved without using depth image.

Without depth input, it seems difficult for the model to estimate the distance to the target object and decide when to stop. Could you please clarify how the stopping condition is implemented in this case?

Thank you for your help.