Xiaohan Mao

The 'configs/detection/cont-det3d_8xb1_embodiedscan-3d-284class-9dof.py' and 'cont-3ddet.pth' you used are correct. Could you please provide more information about the crash in nms_filter?

From the places you changed, I understand that the `results` are not correct, even its length. I can't reproduce this problem but I guess it may be due to incorrect...

> Does the demo in your repository only support 3d detection task for now? If not, then how could we test the capabilities of occupancy and visual grouding tasks in...

Can you please describe the problem in more detail? Do you mean CUDA OOM?

That's interesting, the entire Embodiedscan dataset only takes up about 300G. Are you sure there are no other programs taking RAM up?

We provide `visible_instance_ids` and `visible_occupancy_masks` for each image. It's easy to construct Monocular setting using these masks.

@chanhee-luke Following the [guidance](https://github.com/OpenRobotLab/EmbodiedScan/blob/main/data/README.md), you can find `visible_occupancy.pkl` for each scene. It is a list of `visible_occupancy_annotation` which contains the `img_path` and corresponding `visible_occupancy`.

It seems that there's a misunderstanding about the definition of `occupancy`. Following [TPVFormer](https://github.com/wzzheng/TPVFormer), our `occupancy` is the semantic labels of dense voxels in 3D space.

I notice that `mmengine` warns "Failed to import `None.registry` make sure the registry.py exists in `None` package". Maybe it is not installed correctly?

@Fan123456-c Our `install.py` script will try to install MinkowskiEngine by pip. If it fails, you can follow the the [installation guidance](https://github.com/NVIDIA/MinkowskiEngine?tab=readme-ov-file#installation) of MinkowskiEngine and try other approaches.