Toward Versatile Small Object Detection with Temporal-YOLOv8

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Accurate detection of small objects is a persistent challenge in automated object detection using deep learning.
• Existing deep learning detectors often overlook valuable temporal information in videos, crucial for low signal-to-noise scenarios.
• Current datasets for small object detection are frequently task-specific, lack diversity, and suffer from poor annotations.

Purpose of the Study:

• To develop a versatile deep learning pipeline for accurate small object detection.
• To address limitations in current methods, including feature distinctiveness, temporal information utilization, and dataset quality.
• To improve upon existing object detection architectures like YOLOv8 for small object recognition.

Main Methods:

• Leveraging temporal context from video data to enhance feature representation.
• Implementing data augmentation techniques specifically designed for small objects.
• Utilizing an in-house dataset comprising diverse civilian and military objects for model training and validation.
• Comparing performance against baseline YOLOv8 and models trained on public datasets.

Main Results:

• Achieved a substantial performance increase in YOLOv8, raising mean Average Precision (mAP) from 0.465 to 0.839.
• Demonstrated the effectiveness of incorporating temporal information and tailored data augmentations.
• Showcased the superiority of a model trained on a diverse, carefully curated dataset over environment-specific models.
• Validated the model's capability for accurate small object detection in varied environments.

Conclusions:

• The proposed deep learning pipeline significantly improves small object detection accuracy by utilizing temporal context and specialized augmentations.
• A diverse and well-annotated dataset is critical for developing robust small object detectors.
• The enhanced YOLOv8 architecture offers a fast and accurate solution for detecting small objects across a wide range of applications.