On Salience-Sensitive Sign Classification in Autonomous Vehicle Path Planning: Experimental Explorations with a Novel Dataset

Greer, Ross; Isa, J. M.; Deo, Nachiket; Rangesh, Akshay; Trivedi, Mohan M. · 2022 · Unknown

DOI: 10.1109/wacvw54805.2022.00070

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 challenge of determining traffic sign salience in autonomous vehicle path planning. While existing methods effectively detect and recognize traffic signs, they fail to distinguish which signs are relevant to the ego vehicle’s specific path versus those intended for other lanes or agents. The authors argue that treating all detected signs as equally important can lead to unsafe or illogical maneuvers, particularly in complex urban environments with excessive signage. To address this, the paper introduces the concept of "sign salience," defined as whether a sign provides an instruction intended for the ego vehicle before its next decision point, independent of other agents. To facilitate research in this area, the authors present the LAVA (LISA Amazon-MLSL Vehicle Attributes) dataset, which contains 14,112 samples of traffic signs with bounding boxes, fine-grained category labels, and binary salience labels. The dataset also includes auxiliary metadata such as roadway type and the ego vehicle’s planned maneuver. The authors automatically classify road types (e.g., highway, intersection, residential) using global coordinates, object detection, and temporal context from 10-second video clips. They also classify maneuvers (e.g., forward, stop, turn) based on speed and yaw data. The study evaluates several convolutional neural network (CNN) models to predict sign salience using cropped sign images. A baseline "On-Right" classifier, which assumes salience for signs on the right side of the image, achieved 66.5% accuracy. A ResNet50 model improved this to 74.2%. The authors then tested augmentations by appending one-hot encoded vectors for road type, image coordinates (bounding box location and size), and planned maneuver to the CNN features. The results showed that augmenting with maneuver information yielded the highest accuracy at 76.0%, while road type and coordinate augmentations provided smaller or negligible improvements. Combining all features resulted in 73.6% accuracy, suggesting that maneuver context is the most critical factor for determining salience in this dataset. The significance of this work lies in providing a framework for autonomous vehicles to prioritize regulatory information relevant to their immediate path. By identifying salient signs, downstream control modules can better weight detection errors and make safer, more explainable decisions. The authors conclude that sign salience is a crucial property for safety-critical systems and suggest future work include converting salience to a scalar value and expanding the dataset to improve model robustness.

Key finding

Augmenting a convolutional neural network with information about the ego-vehicle's planned maneuver achieves the highest accuracy of 76% in predicting traffic sign salience.

Methodology

dataset

Sample size: 14112

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.

• signage environment