Field Network-A New Method to Detect Directional Object

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Object detection is crucial in computer vision but faces challenges in accurately describing object shapes.
• Existing methods often rely on Region Proposal Networks (RPN) and anchors, limiting precise shape representation.

Purpose of the Study:

• To propose a new object detection method, Field Network (FN) and Region Fitting Algorithm (RFA), that overcomes limitations of anchor-based approaches.
• To enhance object detection accuracy and shape description using a novel 'Center Field' concept.

Main Methods:

• Introduced the 'Field' concept representing object intensity and probability within an area.
• Developed 'Center Field' to indicate pixel proximity to object centers, abandoning anchors and ROI technologies.
• Utilized an Elliptic Field with normal distribution and RFA for object fitting and direction prediction.

Main Results:

• The proposed Field Network (FN) and Region Fitting Algorithm (RFA) demonstrated improved performance over baseline systems.
• Achieved competitive results on DOTA, MS COCO, and PASCAL VOC datasets.
• The model is efficient with a smaller size (73 M).

Conclusions:

• The Field Network and Region Fitting Algorithm offer a more accurate and efficient approach to object detection.
• The novel 'Field' concept effectively captures object shape and location information.
• This method presents a simpler, smaller, yet highly competitive alternative to current state-of-the-art object detection systems.