Some questions about dataset

[kimkj38]

Thanks for your great work first and I have some questions about the dataset you released.

1. There are 3 perspective images and panorama of 05/scene01005/159. I want to make point cloud of the scene but I guess that I can't get enough point clouds of the scene because the perspective images are very local. Is there any other information?

2. Unlike perspective image, there's only xyz of camera pose in panorama. How can I get rotation of panorama image?

3. I thought I can get panoptic segmentation GT with your instance.png and semantic.png.( e.g. chair1 & chair2) But I couldn't understand how I can use instance.png because it's just binary information that divide whether it is object or not. Could you let me know how I can utilize instance information you provided?

[bertjiazheng]

Hi,

Thanks for your interest in our dataset. Here are some answers:

1. I am afraid there is no other better option.
2. For panorama images, the direction of the camera is always along the negative y-axis.
3. You should read the instance with the correct format. The instance map is coded with unsigned 16-bit integers.

I hope this helps. Feel free to ask if you have any other questions.

Best, Jia

[kimkj38]

Thanks for your help and I have additional questions.

I hope to make 2D bounding boxes of the panorama images from 3D bounding boxes by projection.

But there're 4 panorama images in scene_00000 and all the box information is included in bbox_3d.json so that I can't match the object and the panorama.

Also, I need the class of bounding box but I can't find it.

Can I solve these problems with your data?

[bertjiazheng]

You can obtain the 3D bounding boxes of the visible objects by associating the "ID" in _bbox3d.json and the values in instance.png.

It is unstraightforward to obtain the semantic information of the bounding box. But you can still obtain it by associating the instance.png and semantic.png.

Best, Jia

[kimkj38]

I understood. Thank you!