A review of small object detection in the deep learning era advancements, challenges and future trends

Engineering notes

Key topics: autonomous driving, object detection. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Small object detection (SOD) is a specialized area of computer vision dedicated to localizing and detecting small objects in image and video sequences. While this has significant application across a wide range of fields, SOD presents some unique challenges, including limited pixels, which is the primary reason for their low resolution in the background, poor feature representation, and highly noisy backgrounds. In recent years, detecting small objects has become essential in computer vision. Very advanced deep-learning strategies are employed to overcome the challenges presented. The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) methodology has guided the review process, classifying and evaluating advancements in SOD within the past 16 years. Anchor-free detectors, multiscale feature fusion methods, attention mechanisms, and transformer-based architectures are among the most interesting techniques that have led to a significant performance improvement by improving feature extraction and handling scale variation. Examples of these emerging applications cover a vast array of scenarios: Such applications are used in autonomous driving, for instance, in pedestrian or cyclist detection, in aerial surveillance to identify vehicles or wildlife from satellite imagery, in healthcare diagnosis, for example, by identifying tumors in such medical imagery, and for security with weapons or highly suspicious activity identification. Real-time and lightweight models that can be deployed at the edge and interact with multimodal data sources for self-supervised training with few labels are trends in SOD research. Research that follows these future trends will uncover improvements that will further evolve these challenging issues.

Links and sources

Comments