Introduction
Semantic segmentation is a cornerstone of computer vision, yet its performance is heavily dependent on the quality of training labels. In domains like medical imaging and remote sensing, obtaining perfect "ground truth" is often impossible due to annotator fatigue or expertise gaps. Existing noise-robust methods typically assume that label errors occur independently at each pixel. However, real-world human error is inherently structured; if a pixel on a boundary is mislabeled, its neighbors are likely mislabeled as well. This paper introduces the ELBO-Computable Correlated Discrete Distribution (ECCD), a framework designed to model these spatial correlations efficiently.
Methodology: The ECCD Framework
The core innovation of this work is the treatment of label noise as a spatially correlated process rather than a random pixel-wise event. The authors propose a hierarchical probabilistic model that introduces a latent Gaussian field to represent the likelihood of label errors.
The Latent Gaussian Field and KMS Matrix
Instead of modeling discrete labels directly, the framework utilizes a continuous latent variable, $\eta$, representing the logit of the label error probability. To capture the spatial dependencies between pixels without incurring the prohibitive $O((HW)^2)$ computational cost of standard covariance matrices, the authors employ the Kac-Murdock-Szegö (KMS) matrix.
As shown in Figure 1, the KMS structure allows the model to account for how a pixel's label is influenced by its neighbors. A key mathematical advantage of the KMS matrix is its sparse inverse (tridiagonal) and analytic determinant, which reduces the complexity of calculating the Evidence Lower Bound (ELBO) to $O(HW)$—linear with the number of pixels. This makes the Bayesian approach feasible for high-resolution images.
Experimental Evaluation
The researchers evaluated ECCD on two distinct datasets characterized by structured noise:
- JSRT (Medical Imaging): Chest X-rays where noise was simulated using morphological operations to mimic human boundary errors. ECCD achieved performance nearly identical to training on clean labels, with only a 0.3% drop in accuracy under moderate noise.
- WHU Building (Remote Sensing): Satellite imagery featuring omission and commission errors. ECCD reached an 87.4% IoU, significantly outperforming standard Cross-Entropy baselines that struggle to distinguish signal from structured noise.
Key Insights and Discussion
The success of ECCD stems from its ability to "discount" suspicious clusters of labels during training. By acknowledging that errors happen in contiguous regions, the model effectively transforms hard, noisy one-hot labels into "soft" posterior probabilities.
However, the method does have limitations. It is computationally more intensive than simple cross-entropy due to the need to optimize posterior parameters for every sample. Additionally, its effectiveness decreases for very small or highly elongated structures where spatial dependency is naturally weaker. Future work aims to extend this KMS-based formulation to 3D volumetric data and temporal sequences, where spatial and temporal correlations are even more pronounced.
Conclusion
By bridging signal processing techniques with deep learning, the ECCD framework provides a principled and computationally efficient way to handle spatially correlated label noise. It demonstrates that modeling the structure of noise, rather than treating it as random interference, is key to building robust segmentation models for real-world applications.