Automatic Discovery and Assessment of Interpretable Systematic Errors in Semantic Segmentation

TL;DR

A novel method is presented for discovering systematic errors in segmentation models using multimodal foundation models to retrieve errors and using conceptual linkage along with erroneous nature to study the systematic nature of these errors.

Abstract

This paper presents a novel method for discovering systematic errors in segmentation models. For instance, a systematic error in the segmentation model can be a sufficiently large number of misclassifications from the model as a parking meter for a target class of pedestrians. With the rapid deployment of these models in critical applications such as autonomous driving, it is vital to detect and interpret these systematic errors. However, the key challenge is automatically discovering such failures on unlabelled data and forming interpretable semantic sub-groups for intervention. For this, we leverage multimodal foundation models to retrieve errors and use conceptual linkage along with erroneous nature to study the systematic nature of these errors. We demonstrate that such errors are present in SOTA segmentation models (UperNet ConvNeXt and UperNet Swin) trained on the Berkeley Deep Drive and benchmark the approach qualitatively and quantitatively, showing its effectiveness by discovering coherent systematic errors for these models. Our work opens up the avenue to model analysis and intervention that have so far been underexplored in semantic segmentation.

Figures (9)

• Figure 1: Our framework begins with the inference of an SSM w.r.t. a particular semantic class $c_j$, "person" in this case. Image regions corresponding to the dense predictions for $c_j$ are extracted as patches and are fed to a multimodal foundation model. This model identifies patches which do not represent $c_j$, i.e., patches which denote precision errors. Finally, these precision errors are utilized by an algorithm in order to reveal systematic error groups denoting a common human-interpretable concept.
• Figure 2: Qualitative assessment of systematic errors in the BDD dataset. For the "person" class, concepts of snow and car are systematically present in precision errors, while for the "bicycle" class, the SSMs systematically err on car and metal parts.
• Figure 3: Qualitative results of systematic error discovery in ACDC. Our framework predicts very few false positives and discerns true negatives accurately, showing that its robustness in detecting interpretable systematic errors.
• Figure 4: Precision, recall, and F1-score metrics for precision error identification for "person" where $a = 40$.
• Figure 5: Precision, recall, and F1-score metrics for precision error identification for "person" where $a = 60$.