search

WeijingShi / Point-GNN

Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud, CVPR 2020.
MIT License
523 stars 114 forks source link

Question about the camera points #62

Open thehasnism opened 3 years ago

thehasnism commented 3 years ago

Hi! First of all, I'm really impressed by your work and I have a question about your code.

input_v = cam_rgb_points.attr[:, [0]]

this was in the training file so what is the camera point and why we are using this as input initial vertex features?

Another question is, what happens if we use global coordinates instead of relative coordinates for the edge feature extraction?

Thank you

WeijingShi commented 3 years ago

Hi @thehasnism, Thank you for the questions.

The attr from cam_rgb_points are [lidar_intensity, r, g, b], the lidar intensity are a part of lidar data so we keep them in the vertex features. As for the cam_rgb_points.xyz, they are lidar points in the camera coordinate frame. We train the model and predict the labels in the camera coordinate frame.

The global coordinates distinguish each individual point. When the object translates for distance, the global coordinates of its points changes while the relative coordinates can remain invariant. So if we use global coordinates, the network needs to handle objects at all possible locations. But if we use relative coordinates, the network sees the same object regardless of the object's location. Hope it helps. Thank you.