CoCAtt: A Cognitive-Conditioned Driver Attention Dataset (Supplementary Material)

Shen, Yuan; Wijayaratne, Niviru; Sriram, Pranav; Hasan, Aamir; Du, Peter; Driggs-Campbell, Katherine; Driggs-Campbell, Katherine · 2022 · arXiv

archive: archived pipeline: cataloged verified

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

Summary

This supplementary material details the methodology, baseline architectures, and experimental results for the CoCAtt (Cognitive-Conditioned Driver Attention) dataset. The work addresses the challenge of modeling driver attention under varying cognitive states and driving modes, specifically comparing manual driving with autopilot-assisted driving. A key motivation is the noisy nature of webcam-based eye-tracking data, which suffers from "center shift" issues and random distribution compared to high-fidelity eye-trackers. To mitigate this, the authors propose a gaze calibration network that employs a coarse-to-fine calibration procedure. The study implements several baseline architectures to predict driver attention. These include an unconditioned baseline (BDD-A), a multi-branch model that selects specific sub-branches based on driver states, and a modified CondConv architecture. The modified CondConv integrates driver state information into the routing function of conditional convolution layers, allowing kernel weights to adapt to specific cognitive conditions. The authors increased the dropout rate to 0.7 to prevent overfitting. For the gaze calibration network, webcam gaze inputs are concatenated with encoded spatial features from the upstream network. The models were trained and evaluated on the CoCAtt dataset, which contains data from both manual and autopilot driving modes. Experimental results reveal significant differences in attention modeling between driving modes. Models trained on autopilot data exhibited worse performance than those trained on manual data, indicating that autopilot attention is more variable and harder to model without prior knowledge of planned actions. However, models trained on autopilot data transferred better to manual driving settings than vice versa, likely because drivers in autopilot mode attend to a wider range of potential cues. Quantitative and qualitative analyses demonstrated that cognitive-conditioned learning improves prediction accuracy. Specifically, intention-conditioned models generated more focused attention maps with lower prediction entropy compared to unconditioned models, particularly when approaching intersections. The significance of this work extends to practical applications in road safety analysis. The authors demonstrate a data-driven approach to estimating accident risk by measuring the gaze behavior difference between distracted and attentive drivers using Earth Mover’s Distance. This method identified high-risk areas, such as roundabouts and intersections, where distracted drivers are more likely to miss critical cues. By providing location-wise distraction risk measurements, the cognitive-conditioned attention model offers a generalized solution for assessing road safety standards and identifying hazardous road segments, moving beyond traditional survey-based or historical data methods.

Key finding

Driver-state-conditioned attention models using modified CondConv layers (with state-dependent routing of expert kernel weights) and a coarse-to-fine webcam-gaze calibration network address the center-shift and noise issues of low-cost webcam eye tracking in driver-attention prediction.

Methodology

dataset

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 discover_arxiv on 2026-05-03 (2 acquisition events logged).

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.

• gaze based attention detection
• distraction detection algorithms
• situational awareness
• attention allocation
• cognitive
• visual

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Methodological Resource: tool software
• Theoretical Contribution: computational model, theory or model