Unified Point Cloud Corruption-Aware Reasoning Engine

Grafika Jati, Martin Molan, Francesco Barchi, Andrea Acquaviva

Keywords: Explainable AI, LiDAR Point Cloud Corruption, Graph Attention Networks, Vision-Language Models, Autonomous Vehicle Perception.

TL;DR: The first framework that enables LiDAR sensors to "explain" their own failures.

Abstract: LiDAR sensors are susceptible to surface contamination such as mud, which compromises reliability, yet existing deep learning models fail to provide interpretable diagnostics for such failures. Furthermore, standard Vision–Language Models (VLMs) lack the geometric priors required to reason over LiDAR-specific sparsity and occlusion artifacts. To address this, we present a framework for natural-language explanations of LiDAR corruption through graph-based reasoning and lightweight visual semantics. The proposed method utilizes a Graph Attention Network (GAT) to detect degradation and estimate severity using node-level attention as a geometric saliency signal. Crucially, the system employs a decoupled RGB stream solely for asynchronous semantic verification, discriminating between physical obstructions and sensor-level contamination without heavy multimodal fusion. These GAT-derived signals are mapped to structured prompts and processed by a ViT–GPT2 backbone to synthesize concise, human-readable explanations. Evaluations demonstrate that the proposed method achieves promising qualitative results in contamination detection while providing transparent reliability assessments without the need for end-to-end multimodal retraining.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 50