Rethinking Robustness in Vision-Language Models for Autonomous Driving

Vision-language models (VLMs) are becoming a staple in the scene understanding toolkits of autonomous vehicles. However, the robustness of these models often gets scrutinized through a narrow lens, mainly focusing on how stable their embeddings are. That's missing the point.

Understanding the Real Threat

In an intriguing study, researchers explored the impact of corruption-induced embedding drift on task-aligned hazard scores using CLIP image-text similarities. They employed controlled corruptions from the BDD100K dataset, a massive collection of road scenes, to evaluate the relationship between embedding drift and decision drift, the latter being the change in hazard score due to perturbations.

What's truly striking is how corruption affects these models differently. Some corruption families tightly couple representation drift with decision drift. Others cause decision instability despite modest changes in embeddings. This highlights a potentially dangerous oversight in current robustness benchmarks, which often don't consider the nuanced effects of different types of corruption.

False Negatives and Alarms

What they're not telling you: the direction of failure matters. The study found that while most corruptions tend to suppress hazard detection, leading to false negatives, occlusions did the opposite by triggering false alarms. This suggests a fundamental flaw in evaluating stability by overall instability rates alone without considering these asymmetric failure modes. So, what good is a stable model if it can't distinguish between a real hazard and a phantom threat?

Time for a New Benchmark

The findings here make a compelling case for revamping how we evaluate robustness in VLMs. Current benchmarks focusing solely on embedding-level perturbations are inadequate. They don't capture the task-aligned stability that real-world applications demand. We need task-specific measures that reflect how models perform in the face of specific corruptions that they'll inevitably encounter on the road.

Color me skeptical, but do we really expect autonomous systems to thrive with such glaring gaps in our evaluation methodology? The truth is, aligning robustness benchmarks with real-world tasks isn't just a nice-to-have. it's imperative if these systems are to be trusted on our streets.