Toward Faithful Segmentation Attribution via Benchmarking and Dual-Evidence Fusion

1. Problem Statement

Current attribution methods for semantic segmentation (identifying which pixels cause a model to predict a specific class) are predominantly evaluated based on visual plausibility (i.e., do the heatmaps look reasonable?). However, the authors argue that visual appeal does not guarantee causal faithfulness. Two critical issues remain unaddressed in existing literature:

1. Lack of Causal Verification: Does occluding the highest-attributed pixels actually reduce the model's confidence in the target region?
2. Off-Target Leakage: Do attribution maps incorrectly assign high importance to pixels outside the target object mask (spatially disjoint areas)?

Existing methods, often variants of Gradient-weighted Class Activation Mapping (Grad-CAM), fail to test these causal dependencies rigorously. There is a need for a dedicated evaluation protocol that moves beyond visual inspection to measure true causal dependence and spatial precision.

2. Methodology

A. The Benchmark Framework

The authors introduce a reproducible benchmark to evaluate segmentation attribution methods across four key axes:

1. Target Deletion Faithfulness (TDD): Measures the drop in the model's confidence for a target class when the top-$k$ attributed pixels within the target mask are occluded. A large drop indicates high faithfulness.
2. Off-Target Leakage (ODD & LeakAbs): Measures the sensitivity of the model to occlusions outside the target mask. High importance assigned to non-target pixels is quantified as "leakage."
3. Perturbation Robustness: Evaluates the stability of the attribution map under input perturbations (noise, brightness, blur, flips).
4. Runtime: Wall-clock time per explanation.

The benchmark was tested on Pascal VOC 2012 and SBD datasets using three pretrained backbones: DeepLabV3-ResNet50, FCN-ResNet50, and LRASPP-MobileNetV3.

B. Dual-Evidence Attribution (DEA)

To demonstrate the benchmark, the authors propose Dual-Evidence Attribution (DEA), a lightweight correction method that fuses two types of evidence:

1. Gradient Evidence (EGA): Element-wise gradient-activation products (similar to Seg-XRes-CAM). This provides high-resolution spatial detail but can be unstable or noisy.
2. Region Intervention Evidence (RIA): A causal signal derived by masking fixed grid regions and measuring the drop in class probability. This provides robust causal support but often lacks fine-grained spatial detail (blocky).

The Fusion Mechanism:
DEA combines these signals using an agreement-weighted fusion formula:
$$A_f = \alpha A_g \odot (1 + \beta A_r) + (1 - \alpha) A_r$$
Where:

• $A_g$ is the gradient map, $A_r$ is the intervention map.
• $\odot$ denotes element-wise multiplication.
• The multiplicative term $(1 + \beta A_r)$ amplifies pixels where both gradient and intervention agree.
• The residual term $(1 - \alpha) A_r$ ensures that even if gradients are unstable, the causally supported evidence from the intervention remains.

3. Key Contributions

1. Reproducible Benchmark: A new evaluation framework for segmentation attribution that quantifies intervention-based faithfulness, off-target leakage, and robustness, addressing gaps left by visual-only evaluations.
2. Dual-Evidence Attribution (DEA): A novel method that fuses gradient and intervention signals. It improves deletion-based faithfulness over pure gradient baselines while retaining better robustness than pure intervention methods.
3. Open Source: Release of all per-sample outputs, aggregation scripts, and code to facilitate independent verification.

4. Experimental Results

The authors compared DEA against three baselines: GPA (Gradient-Pooled), EGA (Element-wise Gradient), and RIA (Region Intervention).

• Faithfulness vs. Stability Trade-off:
  • RIA achieved the highest deletion faithfulness (TDD) but suffered from lower stability and higher computational cost.
  • EGA offered the best stability and speed but lower deletion faithfulness.
  • DEA successfully bridged this gap. It consistently improved deletion faithfulness over both gradient baselines (EGA and GPA) across all runs on both datasets.
• Off-Target Leakage: DEA significantly reduced off-target leakage compared to gradient-only methods, ensuring that attribution credit stayed closer to the target mask.
• Robustness: While RIA was the least stable, DEA recovered substantial robustness, performing nearly as well as EGA in perturbation tests.
• Runtime: DEA is slower than pure gradient methods due to the computational cost of intervention passes, but faster than running full intervention baselines for every pixel.

Qualitative Findings:
DEA suppressed broad, low-confidence context activations common in gradient maps while preserving contiguous object structures. Unlike RIA, which could produce "block-like" over-smoothing, DEA maintained sharper intra-object details.

5. Significance and Conclusion

The paper establishes that visual plausibility is insufficient for evaluating segmentation explainability. The proposed benchmark reveals a hidden trade-off between faithfulness (causal correctness) and stability (robustness to noise) that is invisible to human inspection.

DEA is positioned not as a universal "best" method, but as a practical correction for scenarios where explanation quality is prioritized over millisecond latency. It allows researchers to shift gradient-based attributions toward intervention-validated faithfulness without incurring the full computational and stability costs of pure intervention methods.

Limitations & Future Work:

• Current uncertainty reporting relies on run-level variability rather than formal hypothesis testing.
• The fixed-grid intervention approach can miss thin structures, leading to residual leakage in complex scenes. Future work suggests adaptive intervention schemes and formal significance testing.