RGB and NIR Image Segmentation
Controlled comparison of RGB and multimodal inputs for semantic segmentation.
Overview
This computer-vision project compared image-segmentation models trained with RGB information and combined RGB/near-infrared inputs.
Result
| Input configuration | Test score |
|---|---|
| RGB | 87.01% |
| RGB + NIR | 86.45% |
The RGB-only configuration performed slightly better in the final comparison. The result is a useful reminder that adding a sensor modality does not automatically improve generalization; alignment, noise, architecture, normalization, and sample size all affect whether the additional signal is useful.
Evaluation considerations
- Training, validation, and test images were kept separate.
- Model selection was based on validation performance rather than the final test set.
- Input pipelines were compared under a consistent evaluation protocol.
Limitations
The difference between configurations should be interpreted with uncertainty and repeated-run variability in mind. A stronger conclusion would require repeated training seeds and confidence intervals or a paired comparison across test images.
Tools: Python, deep learning, image preprocessing, RGB/NIR data, semantic segmentation