Weakly supervised object localization with soft guidance and channel erasing for auto labelling in autonomous driving...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Automated driving systems (ADSs) require precise object detection for safety.
• Continuous upgrades are needed to adapt object detectors to diverse environments.
• Massive annotations for training data are a significant bottleneck.

Purpose of the Study:

• To propose a novel Weakly Supervised Object Localization (WSOL) method.
• To enable precise object localization without relying on explicit detection annotations.
• To reduce the burden of manual data annotation for ADS development.

Main Methods:

• Development of a novel Weakly Supervised Object Localization (WSOL) method.
• Introduction of the Soft Guidance Module (SGM) and Channel Erasing Module (CEM).
• Integration of SGM and CEM into a mutually beneficial multi-flow framework.

Main Results:

• The proposed WSOL method achieves precise object localization.
• The method effectively reduces the need for extensive detection annotations.
• Experimental validation on Stanford Cars, ILSVRC 2016, and CUB-200-2011 datasets.

Conclusions:

• The novel WSOL method offers an efficient solution for object localization in ADS.
• The Soft Guidance Module and Channel Erasing Module contribute to improved localization accuracy.
• This approach alleviates the challenge of massive annotation requirements in autonomous driving.