TL;DR
This paper introduces a model-agnostic framework to explain pedestrian detection by measuring the importance of specific body parts rather than individual pixels. By replacing standard binary sampling with a novel Beta-distribution-based approach, the authors achieve stable, robust explanations with significantly fewer model evaluations than traditional methods like KernelSHAP.
What Problem Does This Paper Address?
Deep learning models for object detection, especially in autonomous driving (AD), are often treated as "black boxes." While we know that a model detected a pedestrian, we often don't know why. Explainable AI (XAI) methods like KernelSHAP attempt to solve this by perturbing inputs to see how the output changes.
However, applying these methods to images is computationally expensive. Because images have high dimensionality (thousands of pixels), KernelSHAP requires a massive number of forward passes to reach a stable explanation. Furthermore, pixel-level explanations are often noisy and hard for humans to interpret. This paper seeks to make XAI for pedestrian detection both computationally efficient and semantically meaningful.
Methods
The authors propose a pipeline that shifts the focus from pixels to semantic regions—specifically, body parts.
- Body Part Segmentation: They use a pre-trained model called BodyPix to segment pedestrians into 24 parts. To manage complexity, they define four "abstraction levels," grouping these parts into coarser categories (e.g., Level 3 reduces the 24 parts to just 6: head, torso, arms, and legs).
- Superpixel Masking: To test the relevance of a body part, the framework "hides" it. The authors compared three masking techniques: inpainting (filling the area based on surroundings), neighbor noise, and image-wide noise.
- Detection Quality Metric ($q_p$): Instead of simple classification scores, they use a combined metric of the classification confidence and the Sørensen-Dice Coefficient (DICE) to account for bounding box precision.
- Beta Sampling: This is the core innovation. Standard KernelSHAP uses binary sampling (a part is either 100% there or 0% there). The authors introduce a continuous sampling method using a Beta distribution ($\alpha=0.2, \beta=0.1$). This distribution concentrates samples near 0 and 1 (mimicking the Shapley kernel) but allows for "blended" intermediate states, which helps the underlying linear regression converge much faster.
Datasets
The study utilizes the EuroCity Persons dataset, a large-scale benchmark for person detection in urban traffic scenes. They also reference CityPersons as a standard in the field. For the segmentation component, they rely on the BodyPix model, which provides the semantic masks for the pedestrians in the EuroCity images.
Key Insights and Ideas
- Semantic Granularity Matters: By explaining detections through body parts rather than pixels, the results become immediately actionable for engineers. If a model consistently ignores legs, it might struggle with occluded pedestrians behind cars.
- The "Interpolation" of SHAP: The authors view their Beta sampling not as a replacement for the Shapley formalism, but as a robust interpolation of it. It maintains the theoretical spirit of preferring "all-or-nothing" samples while providing enough "in-between" data to stabilize the math.
- Torso and Face Dominance: Across their experiments, the torso and face regions were consistently the most influential features for the RetinaNet50 model they tested. This suggests that the model relies heavily on the central mass and head shape to confirm a "pedestrian" detection.
Key Results
- Efficiency Gain: The Beta sampling method reached stable relevance scores with as few as 64 samples per pedestrian. In contrast, standard KernelSHAP required significantly more samples to converge, and in some cases, failed to provide stable scores even at higher sample sizes.
- Robustness: The proposed method showed higher robustness to lower sampling sizes, making it feasible to run explainability analyses on large-scale datasets that were previously too computationally "heavy."
- Qualitative Alignment: The relevance maps generated by the Beta method were qualitatively similar to KernelSHAP, confirming that the efficiency gain doesn't come at the cost of explanation accuracy.
Limitations and Future Work
- Segmentation Quality: The framework is heavily dependent on the quality of the initial segmentation. The authors noted that BodyPix struggles with low-resolution pedestrians (those far away in the scene) and lacks instance segmentation, making it difficult to analyze crowded groups.
- Dataset Constraints: Currently, there is a lack of large-scale street scene datasets that include "ground truth" body part annotations.
- Simulation Data: The authors suggest that future work should leverage simulation data (like CARLA or SYNTHIA), where perfect semantic and instance masks are available, allowing for even more detailed attribute analysis (e.g., the relevance of backpacks, bikes, or wheelchairs).
Summary updated