Training on customized Dataset

[chrisxu1995]

Hi,
I generate my customized PBR Dataset from BlenderProc for BOP_challenge, and also generate mask and gt from bop_toolkit. Now I uesd run_pose_training.py with config "bop_ycbv_pbr_coarse". And I got the error :

Exception has occurred: AssertionError
exception: no description
File "/home/u2/cosypose/cosypose/lib3d/mesh_ops.py", line 33, in sample_points
assert n_points <= points.shape[1]
File "/home/u2/cosypose/cosypose/lib3d/rigid_mesh_database.py", line 94, in sample_points
return sample_points(self.points, n_points, deterministic=deterministic)
File "/home/u2/cosypose/cosypose/training/pose_forward_loss.py", line 29, in h_pose
points = meshes.sample_points(cfg.n_points_loss, deterministic=False)
File "/home/u2/cosypose/cosypose/training/train_pose.py", line 320, in train_epoch
loss = h(data=sample, meters=meters_train)
File "/home/u2/cosypose/cosypose/training/train_pose.py", line 353, in train_pose
train_epoch()
File "/home/u2/cosypose/run_pose_training.py", line 209, in
train_pose(cfg)

And n_points=2600, points.shape[1]=1148.

What is the reason for this error. what do the n_points und points mean?

[yupei-git]

The problem may lie in the fact that ycbv's objects are relatively large and your objects are small. You can try changing cfg.n_points_loss in run_pose_training.py to a smaller number. In my scenario, I used 500 instead of 2600.

[hannes56a]

Hi @chrisxu1995
were you able to train on your own customized dataset? I have some problems, and i think the reason for my problems are in my 3d data. Do you know any condition for builing own 3d Models for rendering/training?
If i use my own model for training i have problems, if i use a model from "TLESS" and do all the rest on my own (rendering, training), it works...
I have another issue open related to this problem: #64

Do you build also a model on your own? Perhaps you have some hints, tips, ideas?