How to evaluate using the PartObjaverse-Tiny dataset

[zeyu659]

Hello,

Thank you for your work and release!

I encountered two issues while trying to evaluate using the PartObjaverse-Tiny dataset:

1. Do the 200 .glb files in the dataset each need to go through separate Train and Test steps before evaluation? I saw your previous response in https://github.com/Pointcept/SAMPart3D/issues/7, which indicates that only one object can be trained at a time. Is there any update on a method to train them collectively?

2. When using the evaluation scripts eval_semantic.py, eval_instancy.py, and eval_part.py in PartObjaverse-Tiny/eval, could you please provide more specific explanations for meta_path, pred_sem_path, and pred_ins_path?
   During the Train and Test processes, I only see files like
   exp/sampart3d/knight/results/last/mesh_0.0.npy,
   exp/sampart3d/knight/vis_pcd/last/0.0.ply, and
   exp/sampart3d/knight/vis_pcd/last/mesh_0.0.ply,
   which are segmentation results at different scales. However, I don't see the .npy files required by the line pred_sem = np.load(join(pred_sem_path, f"{uid}.npy")) in def eval_all_shape_mean_iou--eval_semantic.py.

[yhyang-myron]

Hi, Thanks for your interest in our work!

1. Yes, only one object can be trained at a time.
2. meta_path: PartObjaverse-Tiny_semantic.json
   pred_ins_path: the folder containing the files like "mesh_0.0.npy", "mesh_0.5.npy"...
   pred_sem_path: need to obtain the semantic prediction first via llm or other models.