Daniel Seichter

Yes, the normal task is working - we did some experiments on Hypersim and NYUv2. However, for NYUv2, the ground truth is no longer available. Training on Hypersim should work....

Class 0 is the void class that indicates _an invalid pixel_ or _a missing annotation_ (for the datasets we consider). Thus, pixels of the void class should be ignored during...

Pretrained weights from PyTorch are loaded automatically for ResNet50 unless you pass the '--no-pretrained-backbone' argument to main.py (see [HERE](https://github.com/TUI-NICR/EMSANet/blob/main/emsanet/args.py#L139)). Loading is done [HERE](https://github.com/TUI-NICR/nicr-multitask-scene-analysis/blob/6a618cf89fff716b8059c0d81a7ac207ef16d4a3/nicr_mt_scene_analysis/model/backbone/resnet.py#L263). However, note that the weights are from...

Yes, of course, the application to Cityscapes is possible. However, we are currently focusing on indoor scene analysis only and did not spend time with preparing Cityscapes for panoptic segmentation....

See our [nicr-scene-analysis-datasets package](https://github.com/TUI-NICR/nicr-scene-analysis-datasets) for further details.

'valid_semantic_miou' is before and 'valid_panoptic_deeplab_semantic_miou' is after merging semantic and instance predictions in order to derive the panoptic prediction. Merging both puts more focus on instances and assigns a single...

Once you have prepared the dataset, you can pass a subsample parameter to the dataset class. If you follow our prepare instructions, every n-th image in each trajectory with n...

For NYUv2: ```bash python main.py \ --tasks semantic normal scene instance orientation \ --enable-panoptic \ --results-basepath /some/path \ --validation-skip 0.95 \ --checkpointing-skip 0.95 \ --checkpointing-metrics valid_semantic_miou bacc panoptic_deeplab_semantic_miou panoptic_all_deeplab_pq panoptic_all_with_gt_deeplab_pq...

Yes. However, note that the actual number of iterations also depends on the specified subset parameter. Even with a random subset of 0.2 or 0.3 per epoch, training on an...

This should not happen. I will run a test training to double-check this.