Incorrect labels in the 2D segmentation dataset

[vobecant]

Dear authors,

I have found some label inconsistencies in the 2D segmentation dataset. For example in the frame 141 (zero-based) of segment-8956556778987472864_3404_790_3424_790_with_camera_labels, for the label 3 (should be truck based on https://github.com/waymo-research/waymo-open-dataset/blob/17f070076dad149766357b31e25d27cf8b5da6ac/waymo_open_dataset/protos/camera_segmentation.proto#L31) there is this:

You can see that the truck on the right is labeled correctly, while the areas on the left do not contain a truck and are labeled incorrectly.

Can you please let us know how often this happens in the dataset?

Thank you very much in advance.

[alexzzhu]

Thanks for calling this out, this kind of error is definitely concerning. However, I can't seem to replicate it on my end. Are you using the functions in utils/camera_segmentation_utils.py to read the dataset? I looked at the data using the colab in tutorial/tutorial_2d_pvps.ipynb and this is what I'm seeing for the semantics on our end:

[xcyan]

@vobecant Hey, do you still face the issue on this specific sequence? I can confirm the labels (frame 141, front camera, with pose_timestamp = 1513450835.6590457 ) are correct in the latest release.

I noticed your issue is related to #634, where frame decoding function or visualization function on your end might be incorrect. Please check the 2d pvps tutorial for details.

More specifically, the code piece under Read a single panoptic label.