Are all objects equal? Deep spatio-temporal importance prediction in driving videos

作者：

Highlights：

• We study a notion of object relevance, as measured in a spatio-temporal context of driving a vehicle.

• Various spatio-temporal object and scene cues are analyzed for the task of object importance classification.

• Human-centric metrics are employed for evaluating object detection and studying data bias.

• Importance-guided training of object detectors is proposed, showing significant improvement over an importance-agnostic baseline.

摘要

Highlights•We study a notion of object relevance, as measured in a spatio-temporal context of driving a vehicle.•Various spatio-temporal object and scene cues are analyzed for the task of object importance classification.•Human-centric metrics are employed for evaluating object detection and studying data bias.•Importance-guided training of object detectors is proposed, showing significant improvement over an importance-agnostic baseline.

论文关键词：Spatio-temporal object analysis,Vision-based behavior analysis,Intelligent and automated vehicles,Human-centric artificial intelligence,Contextual robotics,Driver perception modeling,Object detection

论文评审过程：Received 16 March 2016, Revised 8 July 2016, Accepted 24 August 2016, Available online 30 September 2016, Version of Record 24 December 2016.

论文官网地址：https://doi.org/10.1016/j.patcog.2016.08.029