Daniel Seichter

Ok, I did some test trainings and was able to almost reproduce the reported results in a more recent environment: ``` task: ['semantic', 'scene', 'instance', 'orientation'] task_weighting: [1.0, 0.25, 3.0,...

It looks like you created the dataset with an earlier version of the dataset package. However, for EMSANet, the creation meta file is not necessary, so you can simply comment...

Currently, we do not support mulit-gpu training. Luckily, we have access to A100 GPUs with 40GB/80GB, which have enough VRAM for training with batch size 8.

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...

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...

It is just the input image overlayed with the prediction and some small alpha value. You can easily do it in your pipeline. The function for alpha blending both images...