ESOD: Efficient Small Object Detection on High-Resolution Images.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Enlarging images enhances small object detection but is computationally expensive.
• Current methods for handling high-resolution images in object detection are inefficient.
• Redundant feature extraction on background areas wastes computational resources.

Purpose of the Study:

• To develop an efficient and effective approach for small object detection in high-resolution images.
• To reduce the computational and GPU memory costs associated with image enlargement for object detection.
• To propose a generic framework applicable to various deep learning-based detectors.

Main Methods:

• Reusing the detector's backbone for feature-level object-seeking and patch-slicing.
• Implementing a sparse detection head for efficient processing.
• Integrating the approach into both Convolutional Neural Network (CNN) and Vision Transformer (ViT) based detectors.

Main Results:

• The proposed Efficient Small Object Detection (ESOD) framework significantly reduces computation and memory usage.
• Achieved superior performance on high-resolution inputs (e.g., 1080P).
• Consistently surpassed state-of-the-art (SOTA) detectors, showing up to 8% gains in Average Precision (AP) on datasets like VisDrone, UAVDT, and TinyPerson.

Conclusions:

• The ESOD approach offers an efficient solution for small object detection, overcoming the limitations of simple image enlargement.
• The method demonstrates significant computational savings and performance improvements.
• ESOD is a versatile framework that enhances the capabilities of existing object detection models for small object recognition.