Impact of ISP Tuning on Object Detection.

Area of Science:

• Computer Vision
• Autonomous Systems
• Machine Learning

Background:

• Camera-based object detection is crucial for advanced driver assistance systems (ADAS) and autonomous vehicles.
• Image signal processors (ISPs) are tuned for human perception, with subjective parameter selection impacting image quality.
• The performance implications of ISP parameter variations on deep learning object detection models remain unclear.

Purpose of the Study:

• To analyze the impact of image signal processor (ISP) parameter variations on object detection models used in ADAS.
• To investigate the sensitivity of popular object detection models to changes in eight key ISP blocks.
• To propose a method for improving object detection model robustness against ISP variations.

Main Methods:

• Evaluated 14 object detection models across PASCALRAW and a custom ADAS raw dataset.
• Systematically varied parameters in eight ISP blocks: demosaicing, gamma, denoising, edge enhancement, local tone mapping, saturation, contrast, and hue angle.
• Developed a novel data augmentation methodology based on ISP parameter variations.

Main Results:

• Deviations from default ISP settings generally degrade object detection performance.
• Different object detection models exhibit varying degrees of sensitivity to ISP parameter changes.
• The proposed ISP variation data augmentation technique enhances model robustness.

Conclusions:

• ISP parameter tuning significantly affects camera-based object detection for ADAS.
• Understanding model sensitivity to ISP variations is critical for reliable autonomous systems.
• Data augmentation using ISP variations offers a promising approach to improve the robustness of object detection models.