Aligned Matching: Improving Small Object Detection in SSD.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Generic object detection models, like Single Shot MultiBox Detector (SSD), often struggle with precise detection of small objects.
• The Intersection over Union (IoU)-based matching strategy in SSD can lead to inefficient training for small objects due to suboptimal default box and ground truth matches.

Purpose of the Study:

• To enhance the performance of SSD in detecting small objects.
• To propose a novel matching strategy that improves training efficiency for small objects.

Main Methods:

• Introduced an 'aligned matching' strategy that incorporates aspect ratios and center-point distance alongside IoU for default box matching.
• Evaluated the proposed strategy on the TT100K and Pascal VOC datasets.

Main Results:

• The aligned matching strategy significantly improved the detection of small objects.
• Performance on larger objects was maintained, and no additional parameters were required.

Conclusions:

• The proposed aligned matching strategy effectively addresses the limitations of IoU-based matching for small object detection.
• This approach offers a more balanced and efficient object detection performance across various object sizes.