Cityscapes Visualization

[drgnho]

Dear author,

I conducted training using the Cityscapes dataset and attempted to visualize the results using the "inference_samples.py" script. However, I encountered an issue with the instance segmentation output, wherein a single object appeared to be split into multiple distinct objects. This problem also had an impact on the semantic (panoptic) segmentation result.

How can I properly visualize the instance segmentation for cityscapes?

Thank you so much for your time.

[danielS91]

All foreground pixels (pixels that belong to thing semantic classes) are assigned to instance centers. However, your semantic segmentation seems to be pretty bad - there is a lot of noise in the lower half of the image. I guess there is something wrong. So:

• which input resolution did you use?
• did you change semantic class weighing to linear? (very important for cityscapes - see ESANet parameters here)
• appm as context module might further help with variying input resolutions

To further reduce bad instance assignments, you can have a look at the --instance-offset-distance-threshold parameter here. Especially for any real-world application, it might to useful to assign void.

[danielS91]

However, panoptic instances in your example image look stange, there is something wrong - there is only one center for the center car, but the assignment is splitted in muliple instances.

I can remember a similar output, when we messed up undoing offset normalizing in nicr_mt_scene_analysis/model/postprocessing/instance.py