search

vincentcartillier / Semantic-MapNet

73 stars 11 forks source link

Recompute the GT semantic maps #11

Closed Abundant-me closed 2 years ago

Abundant-me commented 2 years ago

Hi, vincentcartillier. Thanks for your intelligent works.

I want to recompute the point clouds and GT topdown semantic maps since the download link does not contain all the scenes in mp3d.

Here is what I do. I first run data/build_point_cloud_from_mesh.py and then run compute_GT_topdown_semantic_maps/build_semmap_from_obj_point_cloud.py. Here is the problem I meet. In the first step, it generate '.ply' files. In the second step, it tries to load '.h5' files at line 24 and it can not load the generated '.ply' files successfully.

Do I miss anything? Hope your reply. Thanks!

vincentcartillier commented 2 years ago

Hi, Thanks for reaching out. I have added a script that computes the semantic point cloud and saves it as a .h5 file: data/build_point_cloud_from_mesh_h5.py

Try this out and let me know if that works.

Abundant-me commented 2 years ago

Thanks for your reply.

The new script does work. But I meet other problem as follows:

When I run compute_GT_topdown_semantic_maps/build_semmap_from_obj_point_cloud.py, it raise an IndexError at line 119 and 122: IndexError: index 3213656 is out of bounds for dimension 0 with size 3213656 I guess the index should subtract 1 like flat_map_instance[m.view(-1)] = obj_ids[argmax_flat_spatial_map[m] - 1] for line 119, the same for 122.

When I try the modification as above, I can generate semantic map successfully. But the map does not contain the background class, i.e., map_semantic does not contain value 0 and map_instance does not contain value -1. Maybe my modification is wrong?

vincentcartillier commented 2 years ago

The generated ground-truth maps should contain background information. Any pixel that does not represent an object of interest should have a value of 0 in the map_semantic and -1 in the map_instance.

What version of torch_scatter are you using? (SMNet works with the 1.4.0 version)

Abundant-me commented 2 years ago

The version of torch_scatter that I use is 2.0.5. It is higher than what SMNet needs. Should I have to change the version?

vincentcartillier commented 2 years ago

Yes that would do it. torch_scatter >2.0 has major modifications. You should switch to 1.4.0.

Abundant-me commented 2 years ago

I change the version of torch_scatter to 1.4.0 and it figures out my problem.

Thanks a lot!