What are the two common output/loss configurations for binary semantic segmentation?
Computer VisionLevel: IntermediateBinary segmentation logits and losses
What are the two common output/loss configurations for binary semantic segmentation?
Hints
- How many logits are needed to represent two complementary classes?
- Check the agreement between loss, activation and decision rule.
Answer
One uses two logits per pixel with softmax cross-entropy and argmax; another uses one foreground logit with a sigmoid-based binary loss and a threshold. Dataset classes still distinguish foreground and background. Match channel counts, target encoding and loss settings, and tune the threshold on validation if needed. A repository default threshold is not universally optimal.
Source and licence: OpenMMLab — MMSegmentation FAQ · Apache-2.0 · answer: Quantoros editorial