KeyError: 'fpbpn’ when training with PointNet as the image feature extractor

[AnarchistKnight]

I switched from ResNet18 to PointNet, as said in your paper that PointNet better captures floor boundary. Besides, as presented in the paper, DDPM+PointNet has lower KL-divergence than DDPM+ResNet, which indicates that PointNet might help in fitting the underlying probability distribution. I was curious how mush PointNet helps in MiDiffusion, so I simply switched to PointNet. Unfortunately, KeyError: 'fpbpn' occured at the line 47 of networks\diffusion_scene_layout_mixed.py room_feature = sample_params["fpbpn"]

May I ask if I missed any procedure to preprocess the data so as to train with PointNet as the image feature extractor?

[AnarchistKnight]

And I saw the room mask is mapped into a 64 dimensional vector embedding. To me, saving the room boundary polygon coordinates into a \R^{64} vector is much more simple and straight.

[Xmy1120]

And I saw the room mask is mapped into a 64 dimensional vector embedding. To me, saving the room boundary polygon coordinates into a \R^{64} vector is much more simple and straight.

Have you solved this problem? I have the same problem

[AnarchistKnight]

And I saw the room mask is mapped into a 64 dimensional vector embedding. To me, saving the room boundary polygon coordinates into a \R^{64} vector is much more simple and straight.

Have you solved this problem? I have the same problem

well，I suggest you just use resnet18

[SiyiHu]

Most likely you are missing the step to convert the room mask image to fpbpn (denoted as floor_plan_boundary_points_normals in ThreedFront repo). Please refer to this step.

This script first computes floor_plan_ordered_corners, which are ordered corner coordinates of each floor plan, and then samples a fixed number (default:256) of boundary points to ensure the floor plan features are of the same size. The number of corners can have a pretty big range due to curvy floor boundaries.