Unexpected class called 'background' is always inserted in semantic segmentation task

[supersoob]

Describe the bug
Unexpected class called 'background' is always inserted in semantic segmentation task when training with Common Semantic Segmentation data format for OTX.
This class is never specified in dataset_meta.json but can observe that some pixels are predicted to the background class.

Steps to Reproduce

1. Prepare custom data with dataset_meta.json which does not include the "background" class.
   
2. Run 'otx train' command specifying the dataset path from step 1 in data.yaml or --train-data-roots/--val-data-roots
3. You can see evaluation results from the training log includes unexpected class called 'background' label.
   

Environment:

• OS: Linux Ubuntu 20.04.3
• Framework version: PyTorch 1.13.1
• Python version: Python 3.9.13 but can observe this issue in 3.8 and 3.10 as well
• OpenVINO version:
• CUDA/cuDNN version:
• GPU model and memory: