Hot Anchors: A Heuristic Anchors Sampling Method in RCNN-Based Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detection models generate numerous candidate boxes, most lacking valid objects, creating computational burden and sample imbalance.
• This imbalance between object and non-object samples impedes the performance of current detection algorithms.

Purpose of the Study:

• To propose a new heuristic sampling method for generating candidate boxes in two-stage detection algorithms.
• To improve the detection accuracy and efficiency of existing two-stage object detection methods.

Main Methods:

• Developed a novel heuristic sampling strategy to generate more relevant candidate boxes.
• Applied the method to current two-stage detection algorithms, demonstrating its general applicability.
• Conducted experiments on the COCO dataset to evaluate performance.

Main Results:

• The proposed heuristic sampling method significantly increased detection accuracy compared to baseline models.
• The method also led to notable improvements in computational efficiency.
• Experiments confirmed the general applicability across different two-stage detection architectures.

Conclusions:

• The new heuristic sampling method effectively addresses the challenges of candidate box generation in object detection.
• This approach offers a promising solution for enhancing both the accuracy and efficiency of two-stage detection systems.
• The method is broadly applicable and demonstrates significant performance gains on benchmark datasets.