The Sim-to-Real Gap in Indoor Segmentation
Mohammed Faisal Parvez · Hyderabad, IN
- Abstract
- Do synthetic validation scores predict real-world performance?
- Method
- SegFormer-B0 fine-tuned on a 1,458-image labeled real indoor dataset.
- Key Result
- 0.676 mIoU on real-frame validation — while a synthetic-trained counterpart scored 0.988 val mIoU yet failed on real footage.
- Implication
- Synthetic metrics can wildly overstate deployability.