Projection process

Thanks for sharing your nice work!

However, I have a question about the projection process. I assume the point clouds are fisrtly normalized into [-1,1] or [-0.5, 0.5] in your method. Therefore, given normalized point clouds M, how do you project them on the 2D plane using camera parameters shown in equation 1?

I would appreciate it if you coud give me some hints.

Thank you for your question.
Our model generates a standardized point cloud M with the shape of [n, 3] from the world coordinate system. In the projection process, we first use the camera parameters to rotate the point cloud M to the camera coordinate system to obtain M '[n, 3], then take the x-axis and Y-axis of M' to obtain the two-dimensional projection point set P [n, 2], and finally scale the point set P to the image resolution, which is our projection process.

Thanks for your explanation.

So the generated M is located at the world coordinate system. However, the statement in the paper is that M is in an object centered coordinate system. Is there any transformation?
Is it right to scale the point set P (projections) to the mask resoultion of the object?

world coordinate system is the same as the object-centered coordinate system.
Scaling is just to visualize 2D points onto the silhouette. No scaling during training.

Thanks!

Hi, It also seems to be scaled to the mask resolution during the training phase.

2D_projection_matching/2Dpm/models/models.py

Line 283 in 66d6188

outputs['test_o'] = outputs['coord'] * scale