Issues about KITTI-STEP dataset

[HarborYuan]

Hi @markweberdev ,

Thanks for your work on deeplab2 and the STEP paper.

We are trying to follow the KITTI-STEP dataset in our own project. Howver, we found that for the class ids [12,14,15,16,17,18], the instance id annotations are 0. We acknowldge that they do not have the sequence-consistent annotations, but our question is that:

1. does the curent annotation means that each object of the above class labels is unique in the specific image?
2. if 1 is false, do I need to consider [12,14,15,16,17,18] as thing?

Thansk again.

[markweberdev]

Hi HarborYuan,

Indeed, only the classes 11 and 13 have track annotations (see here: https://github.com/google-research/deeplab2/blob/main/data/dataset.py#L162).

I'm not sure I understand the question right. There is no guarantee that for example, an image only contains one bicycle. However, in case there are two bicycles these would not have different track IDs.

You only need to consider 11 and 13 as things, all other classes are considered to be stuff.

Please let me know, if that answers your question.

[HarborYuan]

Hi,

Thanks for your help. I think you have solved my question.
I just treated the annotations as the cityscapes annotations and thought the thing-stuff definition is the same.

Thanks.

[ruili3]

Hi HarborYuan,

Indeed, only the classes 11 and 13 have track annotations (see here: https://github.com/google-research/deeplab2/blob/main/data/dataset.py#L162).

I'm not sure I understand the question right. There is no guarantee that for example, an image only contains one bicycle. However, in case there are two bicycles these would not have different track IDs.

You only need to consider 11 and 13 as things, all other classes are considered to be stuff.

Please let me know, if that answers your question.

Hi Mark,

I have a subsequent question regarding KITTI STEP. Does the instance annotation ID consistent with the initial instance ID used in the KITTI tracking dataset, considering they are using the same data splits?

Thank you!