Proposal-Free Fully Convolutional Network: Object Detection Based on a Box Map.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Region proposal-based detectors (e.g., Faster R-CNNs) are effective but computationally intensive.
• Proposal-free methods offer a balance between accuracy and speed, gaining popularity.
• Existing proposal-free methods have limitations in performance and efficiency.

Purpose of the Study:

• To propose a novel proposal-free, fully convolutional network (PF-FCN) for object detection.
• To outperform existing state-of-the-art proposal-free object detection methods.
• To introduce a "box map" generation technique for improved bounding box prediction.

Main Methods:

• Developed a proposal-free, fully convolutional network (PF-FCN) utilizing a single-pass approach.
• Introduced a "box map" generation method based on regression training.
• Designed a channel and spatial contextualized sub-network to learn the "box map".

Main Results:

• PF-FCN achieved state-of-the-art results on benchmark datasets.
• Achieved 89.6% mAP on PASCAL VOC 2012 and 71.7% mAP on MS COCO.
• Outperformed baseline Fully Convolutional One-Stage Detector (FCOS) and other proposal-free detectors.

Conclusions:

• PF-FCN demonstrates significant improvements in object detection accuracy and speed.
• The proposed "box map" approach is effective for generating accurate bounding boxes.
• Proposal-free detectors hold great significance for practical applications and future research in computer vision.