Towards Fine-Grained Interpretability: Counterfactual Explanations for Misclassification with Saliency Partition

Lintong Zhang, Kang Yin, Seong-Whan LeeCVPR 2025 · DOI

Researchers have developed a new framework called FG-VCE (Fine-Grained Visual Contrastive Explanation) that explains exactly why a deep learning model misclassifies similar-looking objects, such as specific bird species. By swapping fine-grained features from a misclassified image with those from correctly classified ones, the system generates "counterfactual" explanations that highlight specific anatomical parts—like a beak or a wing—that led the model astray.

The Challenge of Fine-Grained Interpretability

Standard AI "attribution" methods, such as Grad-CAM, typically produce blurry heatmaps that indicate where a model is looking but fail to explain what it is seeing. In fine-grained tasks—such as distinguishing a Laysan Albatross from a Frigatebird—these heatmaps often look identical regardless of whether the model's prediction was correct or incorrect. They simply highlight "the bird" without identifying the discriminative features.

Standard attribution maps (right) often fail to distinguish between correct and incorrect predictions in fine-grained tasks, merely highlighting the object's general location.

This lack of granularity is a major hurdle for interpretability. If a model misclassifies an image, we need to know which specific local features caused the error and which features would have led to the correct answer. Existing counterfactual methods often rely on generative models (like GANs), which can be slow, difficult to train, or produce "hallucinated" features that do not correspond to real-world data.

The FG-VCE Framework

The proposed FG-VCE framework operates in three main stages to provide high-fidelity, non-generative explanations:

  1. Feature Extraction: The model processes the misclassified image ($I$) and a set of correctly classified reference images ($U$).
  2. Saliency Partition (SP): To avoid the "blurry heatmap" problem, the authors use a set of spatially localized Gaussian kernels to isolate individual feature points. They then calculate an approximate Shapley value for each point—measuring how much the model's prediction changes when that specific local feature is muted. This disentangles the influence of neighboring pixels.
  3. Iterative Counterfactual Generation: Instead of generating a new image from scratch, the framework iteratively replaces the most "problematic" features in the misclassified image with features from the reference set. It selects replacement features that are semantically similar but contribute strongly to the correct class.
The complete FG-VCE pipeline, illustrating feature extraction, saliency partitioning via Shapley values, and iterative counterfactual generation.

By comparing the image before and after these swaps, the framework identifies Invariant Regions (features that consistently led to the wrong choice) and Dominant Regions (the key features that eventually corrected the prediction).

Key Insights and Ideas

  • Fidelity through Real Features: By using real features from existing data samples rather than generating them with a GAN, the explanations remain grounded in the actual data distribution the model was trained on.
  • Precision via Shapley Values: Using Shapley values (a concept from game theory) allows the system to quantify the "fair share" of a prediction's score attributable to a specific wing-tip or eye-patch, providing much higher resolution than standard gradient methods.
  • Efficiency: Unlike many counterfactual methods that require heavy optimization, FG-VCE is relatively lightweight, using less GPU memory and time than several state-of-the-art competitors.

Experimental Results

The researchers evaluated their framework on two gold-standard datasets: CUB-200-2011 (birds) and Stanford Dogs. To measure precision, they introduced the Compact Activation Score ($\xi$).

  • Quantitative Superiority: On the CUB-200-2011 dataset using a ResNet-50 backbone, FG-VCE achieved an Insertion score of 43.02% and a Deletion score of 10.27%, outperforming traditional methods like Grad-CAM and recent contrastive methods like CCE.
  • Visual Clarity: Qualitatively, the maps produced by FG-VCE are significantly more "compact" and focused on specific anatomical parts compared to the broad, overlapping blobs produced by other methods.
  • Speed: FG-VCE generated explanations in roughly 5.2 seconds, nearly three times faster than some PGD-based attack methods (16.9s).

Limitations and Future Work

While highly effective for CNN-based architectures, the current implementation is tailored to convolutional feature maps. Future work will focus on adapting this "Saliency Partition" logic to Transformer architectures, which handle spatial information differently through self-attention mechanisms. Additionally, the method requires a reference set of correctly classified images, which might be a limitation in extremely data-scarce environments.

Summary updated

BibTeX
@inproceedings{zhang2025towards,
  author = {Lintong Zhang and Kang Yin and Seong-Whan Lee},
  title = {{Towards Fine-Grained Interpretability: Counterfactual Explanations for Misclassification with Saliency Partition}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.02797},
}
Figure