Salient Sign Detection In Safe Autonomous Driving: AI Which Reasons Over Full Visual Context

Greer, Ross; Gopalkrishnan, Akshay; Deo, Nachiket; Rangesh, Akshay; Trivedi, Mohan M. · 2023 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2301.05804

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical challenge of detecting traffic signs in autonomous driving, specifically focusing on identifying "salient" signs—those that directly influence the ego vehicle’s immediate decisions. While standard object detection models aim for uniform performance across all signs, the authors argue that errors are inevitable and should be strategically distributed. Missing a non-salient sign (e.g., a speed limit for a distant side street) is less hazardous than missing a salient one (e.g., a stop sign for the current lane). To address this, the study introduces a framework that prioritizes the detection of signs pertinent to the vehicle’s current lane and immediate path, thereby enhancing safety by ensuring the most critical information is captured. The methodology involves three primary components: dataset creation, model selection, and loss function modification. First, the authors constructed the LAVA Salient Signs Dataset, extending the existing LAVA dataset with 31,191 annotated traffic sign instances. Each sign was labeled with a "salience" property based on whether it influences the ego vehicle’s next immediate decision, considering factors like lane position, proximity to intersections, and sign orientation. Second, they employed Deformable DETR, a transformer-based object detection model capable of reasoning over full visual context with reduced computational cost compared to standard DETR. Third, they introduced a novel "Salience-Sensitive Focal Loss." This custom loss function modifies the standard focal loss by applying a higher weight ($\alpha_{ss} = 4$) to the classification error of salient signs, effectively forcing the model to prioritize learning features relevant to these critical objects during training. Experimental results demonstrate that training Deformable DETR with the Salience-Sensitive Focal Loss yields superior performance compared to the baseline model trained with standard loss. The modified model achieved higher recall for both salient signs and all signs combined. Crucially, the performance margin between salient sign recall and overall sign recall was largest for the model using the salience-sensitive loss, indicating that the model successfully shifted its focus to prioritize critical signs without significantly degrading general detection capabilities. The authors attribute this improvement to the loss function guiding the transformer’s attention mechanisms toward image regions containing salient signs, which often co-occur with other sign types, thereby improving feature extraction for the entire scene. The significance of this work lies in its contribution to safe autonomous driving systems by aligning detection priorities with safety-critical requirements. By formalizing sign salience and integrating it into the training process via a specialized loss function, the study provides a method to reduce high-risk detection errors. The creation of the LAVA Salient Signs Dataset also offers the research community a new resource for developing and evaluating salience-aware perception models. The findings suggest that incorporating semantic importance into detection architectures can enhance overall system reliability, ensuring that autonomous vehicles remain aware of the most impactful environmental cues.

Key finding

Training a Deformable DETR model with a Salience-Sensitive Focal Loss function significantly improves the recall of salient traffic signs and overall sign detection performance compared to standard training methods.

Methodology

dataset

Sample size: 31992

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-28.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms