SI-FID boosts stitched image quality scoring by 25% over human correlation
A new metric detects ghosting and misalignment 25% better than existing methods.
Evaluating the quality of stitched images has long been a challenge because existing objective metrics—like PSNR or SSIM—fail to capture human perception of artifacts such as ghosting, misalignment, and color inconsistencies. To bridge this gap, Shengwei Guo and Guobing Sun introduce SI-FID, a noise-aware fine-tuning approach built on top of the Fréchet Inception Distance (FID). Instead of modifying the FID formula, they adapt its underlying feature space using contrastive learning with controlled data perturbations. A pre-trained InceptionV3 encoder is calibrated on both original and artificially distorted samples, making the feature representation more sensitive to the subtle distortions that characterize poor stitching.
The results are significant: on two complementary benchmark datasets, SI-FID achieves over 25% improvement in Spearman rank correlation with human subjective scores compared to conventional metrics. This means the metric aligns much more closely with how people actually perceive stitched image quality. By providing a perceptually consistent and reliable indicator, SI-FID could accelerate progress in panorama generation, virtual reality content creation, and any application where seamless image blending is critical. The technique is lightweight enough to be integrated into existing stitching pipelines as a drop-in evaluation module.
- SI-FID adapts InceptionV3 features via contrastive fine-tuning with data augmentation to detect subtle stitching artifacts.
- Improves Spearman rank correlation with human subjective scores by over 25% across two benchmark datasets.
- Targets specific artifacts like ghosting and misalignment that conventional metrics usually miss.
Why It Matters
More reliable stitched image scoring enables faster iteration in panorama and VR content creation workflows.