LEOD-Net: Learning Line-Encoded Bounding Boxes for Real-Time Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Object detection commonly uses bounding boxes defined by corner points.
• Existing methods often involve complex pipelines for accurate object localization.

Purpose of the Study:

• To propose a simplified, learnable line encoding technique for bounding boxes in object detection.
• To enhance the efficiency and implementation ease of object detection models.

Main Methods:

• Encoding bounding boxes using two primary corner points.
• Employing a lightweight convolutional neural network (CNN) with pixel-shuffle for high-resolution line mask generation.
• Utilizing progressive probabilistic Hough transform for line mask post-processing and refinement.

Main Results:

• Achieved high mean average precision (mAP) on Pascal VOC2007 (78.8%) and MS-COCO2017 (48.1%).
• Demonstrated fast processing speeds (37 ms on VOC, 47 ms on COCO).
• The method proved simpler and easier to implement compared to complex state-of-the-art approaches.

Conclusions:

• The proposed learnable line encoding offers a simple yet effective approach to object detection.
• This method balances high accuracy with computational efficiency.
• It presents a promising alternative to complex object detection pipelines.