Learning Efficient Binarized Object Detectors With Information Compression.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Conventional binarized neural networks for object detection suffer from performance degradation due to information redundancy and limited representational capacity.
• Direct quantization in existing methods leads to numerous false positives and reduced detection accuracy.

Purpose of the Study:

• To develop an efficient binarized neural network learning method (BiDet) for object detection that enhances precision and reduces false positives.
• To introduce AutoBiDet, which automatically adjusts information compression based on input complexity for improved performance.
• To propose class-aware sparse object priors for more effective false positive elimination without compromising recall.

Main Methods:

• Generalizing the information bottleneck (IB) principle to object detection by constraining high-level feature map information and maximizing mutual information.
• Learning sparse object priors to concentrate predictions on informative detections and eliminate false positives.
• Developing AutoBiDet with automatic IB trade-off adjustment for varying input complexities.
• Proposing class-aware sparse object priors to assign different sparsity levels to various object classes.

Main Results:

• BiDet effectively utilizes representational capacity through redundancy removal, enhancing detection precision and alleviating false positives.
• AutoBiDet demonstrates improved performance by dynamically adjusting information compression based on input complexity.
• Class-aware sparse object priors further reduce false positives without a decrease in recall.
• Both BiDet and AutoBiDet significantly outperform existing state-of-the-art binarized object detectors on PASCAL VOC and COCO datasets.

Conclusions:

• BiDet and AutoBiDet represent significant advancements in efficient and accurate binarized object detection.
• The proposed methods effectively address the limitations of conventional binarization techniques.
• These approaches offer a promising direction for developing high-performance, low-resource object detection systems.