iffDetector: Inference-Aware Feature Filtering for Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Modern object detectors often overlook feature optimization during inference.
• Existing methods typically use open-loop feature calculations without feedback mechanisms.

Purpose of the Study:

• To propose a novel feature optimization approach for enhancing object detection.
• To introduce an inference-aware feature filtering (IFF) module for CNN-based detectors.

Main Methods:

• Developed a generic IFF module for closed-loop feature optimization using high-level semantics.
• Integrated the IFF module into existing detectors, creating the iffDetector.
• Analyzed the IFF module's stability using Fourier transform, demonstrating its negative feedback properties.

Main Results:

• The iffDetector consistently outperformed state-of-the-art methods on PASCAL VOC and MS COCO datasets.
• The IFF module demonstrated significant improvements in feature enhancement and background noise suppression.
• The plug-and-play IFF module added minimal computational overhead.

Conclusions:

• The proposed IFF module effectively optimizes features during both training and inference in object detectors.
• The iffDetector offers a stable and efficient solution for improving object detection performance.
• This approach provides a significant advancement in CNN-based object detection technology.