search

carla-simulator / carla

Open-source simulator for autonomous driving research.
http://carla.org
MIT License
11.65k stars 3.75k forks source link

Instance Segmentation - Standard semantic ID only contains subset of objects #5842

Open jokolaco opened 2 years ago

jokolaco commented 2 years ago

According to the documentation " The R channel contains the standard semantic ID." While this is true and working in the semantic segmentation case i noticed in the instance segmentation case that e.g. cars are in the red channels 9,10,11 instead of only the 10th redchannel.

This makes it hard to access the segmented areas because they are not clearly living in one specific channel. This is also the case for the "actor id" tagging.

Input: 586 Mask: 586_semseg

Redchannel: 9 InstanceSegmentation_Mask_r9 10 InstanceSegmentation_Mask0r10 11 InstanceSegmentation_Maskr11

CARLA version: 9.13 and latest night build, standard downloads from github and self compiled on Rocky Linux 8

Platform/OS: Ubuntu 18 and Rocky Linux 8

mozg1 commented 1 year ago

Can the Instance Segmentation sensor finally be looked at for the next release, because it is simply not working according to the documentation. My thesis depends on this, thanks.

JosenJiang commented 1 year ago

Do you save instance segmentation image as JPG format? Please save it as PNG format. JPG is a lossy compression format, and some pixel's value will change when you save images as xx.jpg. PNG is lossless compression format.

stale[bot] commented 1 year ago

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

CaphalorAlb commented 1 year ago

Do you save instance segmentation image as JPG format? Please save it as PNG format. JPG is a lossy compression format, and some pixel's value will change when you save images as xx.jpg. PNG is lossless compression format.

hey, I'm not sure how to do this properly, running into the same issue

this is what i'm using

image_pv_seg.save_to_disk(frame_path_seg + '.png')

CaphalorAlb commented 1 year ago

According to the documentation " The R channel contains the standard semantic ID." While this is true and working in the semantic segmentation case i noticed in the instance segmentation case that e.g. cars are in the red channels 9,10,11 instead of only the 10th redchannel.

This makes it hard to access the segmented areas because they are not clearly living in one specific channel. This is also the case for the "actor id" tagging.

Input: 586 Mask: 586_semseg

Redchannel: 9 InstanceSegmentation_Mask_r9 10 InstanceSegmentation_Mask0r10 11 InstanceSegmentation_Maskr11

CARLA version: 9.13 and latest night build, standard downloads from github and self compiled on Rocky Linux 8

Platform/OS: Ubuntu 18 and Rocky Linux 8

did you ever find a solution for this? running into the same issue and i'm not sure how to work around it

Blyron commented 1 month ago

Due to a contribution from @AreopagX Now the actor id generated by carla matches the instance segmentation one; For docs check; https://carla.readthedocs.io/en/latest/ref_sensors/#instance-segmentation-camera https://carla.readthedocs.io/en/latest/tuto_G_instance_segmentation_sensor/ https://carla.readthedocs.io/en/latest/tuto_D_create_semantic_tags/

Usually the id in red channel is Unreal generated actor id which differs from CARLA's. that's why everyobject can be shown in the red channel.