PaperlayerView exact revision on arXiv

Source-linked AI summary

Small Object Detection: A Comprehensive Survey on Challenges, Techniques and Real-World Applications

Mahya Nikouei, Bita Baroutian, Shahabedin Nabavi, Fateme Taraghi, Atefe Aghaei, Ayoob Sajedi, Mohsen Ebrahimi Moghaddam

arXiv:2503.20516v1cs.CV

TL;DR

Small object detection faces limited resolution, background interference, and class imbalance. This survey reviews deep-learning research published in Q1 journals during 2024–2025. It highlights multi-scale feature extraction, attention mechanisms, and super-resolution as approaches addressing these challenges.

  • Problem

    Small object detection faces limited resolution, background interference, and class imbalance across surveillance, autonomous systems, medical imaging, and remote sensing.

    Source: 8- Conclusion · p. 38

  • Method

    The survey reviews challenges, techniques, datasets, metrics, and applications through Scopus-indexed Q1 journal articles published during 2024–2025 with “Small Object Detection” in their titles.

    Source: 1- Introduction · p. 3

  • Results

    The survey highlights multi-scale feature extraction, attention mechanisms, and super-resolution as approaches addressing small-object detection challenges.

    Source: 8- Conclusion · p. 38

  • Takeaways & Limitations

    Datasets tailored for small object detection and robust evaluation metrics support more accurate benchmarking and model development.

    Source: 8- Conclusion · p. 38

    Popular detectors remain computationally intensive on high-resolution images, increasing resource demands and processing time (J. Liu et al., 2024).

    Source: 3-4- Limitations of Popular Detection… · p. 8

Abstract

from arXiv · show

Small object detection (SOD) is a critical yet challenging task in computer vision, with applications like spanning surveillance, autonomous systems, medical imaging, and remote sensing. Unlike larger objects, small objects contain limited spatial and contextual information, making accurate detection difficult. Challenges such as low resolution, occlusion, background interference, and class imbalance further complicate the problem. This survey provides a comprehensive review of recent advancements in SOD using deep learning, focusing on articles published in Q1 journals during 2024-2025. We analyzed challenges, state-of-the-art techniques, datasets, evaluation metrics, and real-world applications. Recent advancements in deep learning have introduced innovative solutions, including multi-scale feature extraction, Super-Resolution (SR) techniques, attention mechanisms, and transformer-based architectures. Additionally, improvements in data augmentation, synthetic data generation, and transfer learning have addressed data scarcity and domain adaptation issues. Furthermore, emerging trends such as lightweight neural networks, knowledge distillation (KD), and self-supervised learning offer promising directions for improving detection efficiency, particularly in resource-constrained environments like Unmanned Aerial Vehicles (UAV)-based surveillance and edge computing. We also review widely used datasets, along with standard evaluation metrics such as mean Average Precision (mAP) and size-specific AP scores. The survey highlights real-world applications, including traffic monitoring, maritime surveillance, industrial defect detection, and precision agriculture. Finally, we discuss open research challenges and future directions, emphasizing the need for robust domain adaptation techniques, better feature fusion strategies, and real-time performance optimization.

1- Introduction … 4-1- Recent Trends

This survey reviews deep-learning small object detection research from 2024–2025 Q1 journals, examining definitions, detection challenges, and emerging techniques. Recent research emphasizes stronger backbones, feature extraction and fusion, attention, and advanced learning strategies.

1- Introduction

  • The survey targets 2024–2025 Scopus-indexed articles titled with “Small Object Detection” in SJR Q1 journals, reviewing challenges, techniques, datasets, metrics, and applications.Small objects contain limited spatial and contextual information, yet missed detections can undermine safety-critical systems. Applications include autonomous navigation, remote sensing, and detecting small medical abnormalities for early diagnosis (Q. Feng et al., 2023).Source: 1- Introduction · pp. 2–3

2- Definitions and Background / 2-1- Pixel-Based Definitions / 2-2- Relative Size Criteria / 2-3- Categorization of Small Objects

  • Less than 1,024 square pixels defines small objects in MS COCO; relative-size thresholds and tiny/dense categories provide additional definitions.Definitions depend on dataset and application: some studies use less than 1% of image area (B. Liu & Jiang, 2024; L. Ni et al., 2024; Tong & Wu, 2024; Yang et al., 2024). Reported categories include tiny objects up to 4500 square pixels, small objects above the tiny category but below dataset-specific thresholds, and overlapping dense small objects in crowded UAV scenes (C. Chen et al., 2024).Source: 2-1- Pixel-Based Definitions · p. 4, 2- Definitions and Background · p. 4, 2-2- Relative Size Criteria · p. 4

2-4- Challenges Highlighted by Definitions / 3- Challenges in SOD / 3-1- Limited Appearance Information and Occlusion

  • Limited pixels, background blending, scale variability, and occlusion undermine recognition, producing missed or false detections in real-world small object detection.Occlusion compounds feature indistinguishability (Ma et al., 2025; J. Wu et al., 2024), while lighting changes, motion blur, dust, low illumination, and water vapor further degrade accuracy (Aibibu et al., 2023a). These challenges affect surveillance, medical imaging, remote sensing, and industrial inspection, with drone imagery and coal mines illustrating detection failures.Source: 2-4- Challenges Highlighted by… · p. 5, 3- Challenges in SOD · p. 5, 3-1- Limited Appearance Information… · p. 7

3-2- Challenges in Localization and Scale Variation

  • Spatial uncertainty, demanding localization precision, and scale-dependent anchor mismatches make small-object bounding boxes difficult to predict accurately.High anchor-box IoU is difficult to achieve, fine-grained image analysis is computationally intensive, and scale variation can misalign receptive fields (S. Li et al., 2024; S. Chen et al., 2024; Sun et al., 2024; D. Chen et al., 2024; Jing, Zhang, Li, et al., 2024b). Low-resolution coal-mine imagery further degrades detection, while centralized-cloud latency and concept drift impede timely, reliable deployment (J. Wu et al., 2024).Source: 3-2- Challenges in Localization and… · pp. 7–8

3-3- Inefficiency in Feature Learning and Background Interference / 3-4- Limitations of Popular Detection Methods

  • Downsampling removes discriminative details, while FPNs trade shallow-layer localization against deep-layer semantics, limiting small-object feature learning.Complex backgrounds also lower signal-to-noise ratio, obscuring industrial defects and aerial targets (D. Zhao et al., 2024). Detectors designed primarily for larger objects lose fine details and incur high-resolution processing costs (J. Liu et al., 2024); multi-task imbalance limits learning (S. Li et al., 2024), and adapting architectures, training, or losses can be expensive (C. Chen et al., 2024).Source: 3-3- Inefficiency in Feature Learning… · p. 8, 3-4- Limitations of Popular Detection… · pp. 8–9

3-5- High Computational Costs and Hardware Resource Limitations / 3-6- Inconsistent Performance Across Different Scales and Datasets

  • Limited embedded computing constrains real-time deployment, while training–testing scale differences and scarce specialized datasets undermine consistent detection performance.UAV, USV, and coal-mine systems must reconcile high-resolution processing with limited resources, latency requirements, and sometimes few-shot adaptation to new categories. Scale mismatch can exclude objects from anchor/grid training regions, repeated convolution can erase features (Tong & Wu, 2024; L. Zhou et al., 2024), and dense scenes plus insufficient aerial and industrial datasets exacerbate degradation.Source: 3-5- High Computational Costs and… · p. 9, 3-6- Inconsistent Performance Across… · p. 9

3-7- Solutions and Emerging Approaches

  • Motion–visual learning, Multi-Granularity Detection (D. Chen et al., 2024), and enhanced FPNs address complementary small object detection weaknesses.Motion-inspired cross-pattern learning combines motion and visual cues for dynamic environments, while Multi-Granularity Detection addresses sample imbalance and detection across scales. Enhanced FPNs fuse low- and high-level features to narrow the semantic gap while preserving localization information.Source: 3-7- Solutions and Emerging Approaches · pp. 9–10

4- Deep Learning Techniques for SOD / 4-1- Recent Trends

  • 23.1% of reviewed research focuses on optimized backbone architectures, the largest share among the reported small object detection trends.Other shares are attention mechanisms, 18.5%; feature extraction enhancement, 16.9%; feature fusion optimization, 15.4%; advanced learning strategies, 13.8%; and multi-scale hybrid attention and high-resolution feature alignment, 12.3%. Techniques include optimized backbones and enhancement modules, multi-level fusion through FPNs and lightweight integration, channel/spatial/deformable/transformer attention to suppress background interference, and Knowledge Distillation (Nabavi, et al., 2024) and Reinforcement Learning for deployment adaptability.Source: 4-1- Recent Trends · pp. 10–11, Figure 3 · p. 12

4-1-1- Model Optimization and Lightweight Architectures … 4-2- Neural Network Architecture

This section reviews neural network architectures and optimization strategies for small-object detection, emphasizing multi-scale feature extraction, feature fusion, attention, and lightweight design. It also covers anchor-free and transformer models, data augmentation, knowledge distillation, and self-supervised learning to address detection accuracy, computational efficiency, and limited training data.

4-1-1- Model Optimization and Lightweight Architectures

  • Lightweight models use modular design, parameter sharing, and compression to balance detection accuracy with efficient deployment on drones, mobile devices, and edge platforms.Examples include FFEDet (F. Zhao et al., 2024) and KDSMALL (W. Zhou et al., 2024).Source: 4-1-1- Model Optimization and… · p. 12

4-1-2- Feature Fusion and Multi-Scale Detection

  • Multi-scale feature extraction and fusion improve contextual understanding and detection precision through attention, transformers, and large kernels (W. Zhou et al., 2024).Multi-frame fusion and cross-modal approaches, including thermal–visible fusion, support detection in haze, sandstorms, and maritime settings (Z. Chen, Shi, et al., 2024; W. Zhou et al., 2024).Source: 4-1-2- Feature Fusion and Multi-Scale… · p. 12

4-1-3- Anchor-Free and Transformer-Based Approaches

  • Anchor-free detectors directly predict bounding boxes without predefined anchors (S. Li et al., 2024), while transformers capture long-range dependencies and multi-scale features.Hybrid CNN-transformer models combine both frameworks (Z. Zhu et al., 2024). DETR-like methods are explored to address Non-Maximum Suppression (NMS) limitations.Source: 4-1-3- Anchor-Free and… · p. 13

4-1-4- Advanced Data Augmentation Techniques

  • GAN-generated images, hard augmentation, and diffusion models address dataset imbalance and scarce small-object samples while improving generalization and training robustness (Q. Liu et al., 2024).Source: 4-1-4- Advanced Data Augmentation… · p. 13

4-1-5- Integration of Multi-Modal and Multi-Domain Data

  • Combining LiDAR, radar, infrared, and visual features enriches representations and improves detection for industrial inspection and intelligent traffic systems (Jing, Zhang, Liu, et al., 2024).Source: 4-1-5- Integration of Multi-Modal and… · p. 13

4-1-6- Knowledge Distillation and Self-Supervised Learning

  • Knowledge distillation improves SOD efficiency and real-time performance, while self-supervised learning reduces dependence on labeled data (W. Zhou et al., 2024).Self-supervised learning supports unsupervised and semi-supervised training; knowledge distillation is also deployed for lightweight SOD on UAVs.Source: 4-1-6- Knowledge Distillation and… · p. 13

4-2- Neural Network Architecture

  • Neural architectures combine feature extraction, fusion, attention, motion processing, deformable convolution, reinforcement learning, and knowledge distillation to improve detection and reduce computational costs.MDFFAM preserves spatial details through multidirectional feature fusion, while LKSPP expands the receptive field and reduces computational complexity.Source: 4-2- Neural Network Architecture · pp. 14–15
  • Two-stage architectures use low-resolution localization and high-resolution classification to improve accuracy while reducing computational cost.The stages separate localization from classification rather than performing both at high resolution.Source: 4-2- Neural Network Architecture · p. 16
  • Multi-YOLOv8 combines frame fusion, optical flow, BiFormer attention, and WIoU v3 to improve sensitivity to small objects.The infrared moving-object detector extracts motion and spatial features, suppresses background, and includes a dedicated detection layer.Source: 4-2- Neural Network Architecture · p. 16
  • The multi-granularity detector combines deformable-convolution feature extraction with dynamic-threshold sample balancing to address sample imbalance (D. Chen et al., 2024).Deformable convolution provides multilevel features, while dynamic thresholds govern sample balancing.Source: 4-2- Neural Network Architecture · p. 16
  • Multi-scale hybrid attention combines cross-scale similarity extraction, foreground enhancement, and channel-redundancy reduction for UAV small-object detection.MsA extracts cross-scale similarities, FEM enhances foreground features, and MHA and DSCA support attention processing.Source: 4-2- Neural Network Architecture · p. 16
  • HV-YOLOv8 uses HDPConv in its backbone and VOV-GSCSP in its neck to improve feature extraction and enable lightweight fusion.These components reduce complexity and optimize detection accuracy.Source: 4-2- Neural Network Architecture · p. 16
  • MFFSODNet integrates multiscale extraction, BDFPN fusion, and a modified prediction head into YOLOv5 for efficient UAV small-object detection.Multiscale extraction enriches features, while BDFPN enhances fusion and the modified prediction head improves small-object detection.Source: 4-2- Neural Network Architecture · p. 17
  • KDSMALL combines EfficientNet features, CBAM attention, and head-level knowledge distillation to increase accuracy while maintaining efficiency.CBAM applies spatial and channel attention to improve fusion; distillation transfers knowledge from a larger model to a smaller one.Source: 4-2- Neural Network Architecture · p. 17
  • RFRM-CenterNet combines a ResNet50 backbone, multiscale receptive-field refinement, and a three-subnetwork head for precise frequency-hopping signal detection.RFRM enhances multiscale features, while the specialized head targets frequency-hopping signals.Source: 4-2- Neural Network Architecture · p. 17
  • MSFEM uses convolutional branches with different kernel sizes to capture fine-grained, multi-scale features and improve small-object detection.The listed kernel sizes include 1×1, 3×3, and 5×5.Source: 4-2- Neural Network Architecture · p. 17
  • Feature fusion combines shallow detail with deep semantics through BDFPN, while AFFM balances features of small and large objects.BDFPN supports small-object detection in complex scenarios.Source: 4-2- Neural Network Architecture · pp. 17–18
  • Architectural optimization targets multi-scale extraction and fusion alongside real-time efficiency, while knowledge distillation in the head aims to enhance accuracy and reduce computational complexity.The architectural overview explicitly identifies reduced computational complexity as a route to real-time efficiency.Source: 4-2- Neural Network Architecture · pp. 17–18
  • Attention mechanisms improve object differentiation and mitigate background noise; EMA separates objects from backgrounds, while MCSTA combines channel and spatial attention.MCSTA denotes multi-head channel and spatial trans-attention.Source: 4-2- Neural Network Architecture · p. 18
  • Soft Pooling minimizes information loss during feature extraction, while Subspace Attention enhances spatial localization of small objects.These mechanisms address feature preservation and localization, respectively.Source: 4-2- Neural Network Architecture · p. 18
  • MsA extracts cross-scale similarities, FEM enhances foreground features, and DSCA reduces channel redundancy.The described MHA structure incorporates these complementary mechanisms for small-object detection.Source: 4-2- Neural Network Architecture · p. 18
  • FEM enhances foreground features, while MHA and DSCA provide attention processing.Source: 4-2- Neural Network Architecture · p. 18
  • The enhanced YOLOv8 design uses DBB for scalable feature extraction, switching from multiple branches during training to a single branch during inference.The branch change is intended to balance complexity and performance.Source: 4-2- Neural Network Architecture · p. 19
  • Enhanced YOLOv8 combines C2f modules, decoupled detection heads, and an anchor-free structure with BiFPN fusion, query-based localization, and DyHead multi-scale attention.The Q-block supplies the query mechanism for localization.Source: 4-2- Neural Network Architecture · p. 19
  • SONet uses multi-branch dilated convolutions for multi-scale feature extraction, improving performance in power-line inspection.The branches capture objects across multiple scales and help recognize details in complex power-line scenes.Source: 4-2- Neural Network Architecture · p. 19
  • AAFF refines and fuses features to emphasize small power-line defects, while β-CIoU optimizes bounding-box regression.Attention mechanisms direct feature fusion toward critical regions of interest.Source: 4-2- Neural Network Architecture · p. 19

4-3- Clarity and Visual Information Improvement

Improving image clarity helps models extract and use fine-grained details, especially for small or low-resolution objects. The section covers SR, multiscale information, and fusion across neural-network layers (D. Chen et al., 2024).

4-3-1- SR for Image Quality Enhancement

Super-resolution (SR) enhances low-resolution images so detection models can distinguish tiny objects and improve their localization and classification. Deep-learning approaches recover fine details, while complementary enhancement techniques help prevent artifacts.

4-3-1- SR for Image Quality Enhancement

  • CNN-based SR models learn to upscale low-resolution inputs, using contextual information to reconstruct missing high-resolution details (K. Liu et al., 2025).Source: 4-3-1- SR for Image Quality Enhancement · p. 20
  • SR improves small-object localization and classification by enhancing image resolution and revealing minute details otherwise obscured by resolution constraints.Source: 4-3-1- SR for Image Quality Enhancement · p. 20
  • GAN-based SR generates realistic high-resolution images by learning the underlying image-data distribution (Y. Zhao et al., 2024).SR is often combined with denoising and deblurring to avoid artificial artifacts that could degrade model performance.Source: 4-3-1- SR for Image Quality Enhancement · p. 20

4-3-2- Utilizing Multi-Scale Information

Multi-scale information supports small object detection by combining fine-grained details with broader context, while requiring careful balancing of spatial and semantic information.

Utilizing Multi-Scale Information

  • Multi-scale feature extraction captures details and context needed to detect differently sized objects (D. Liao et al., 2025a; Y. Zhao et al., 2024).Features are extracted and processed at multiple scales and levels of abstraction.Source: 4-3-2- Utilizing Multi-Scale Information · p. 20
  • Multi-resolution networks and feature pyramids combine information across scales to recognize both small and large objects.Early-layer details can be combined with deeper, abstract features (F. Feng et al., 2025b; Song et al., 2024; B. Wang et al., 2024).Source: 4-3-2- Utilizing Multi-Scale Information · p. 20
  • Balancing spatial detail and semantic information remains challenging because fine-grained details can be lost across scales.FPN combines high-resolution fine-grained features with coarser context from deeper layers to address this challenge (Z. Chen, Ji, et al., 2024; Gao, Li, et al., 2024; Q. Liu et al., 2024).Source: 4-3-2- Utilizing Multi-Scale Information · p. 21

4-3-3- Fusion of Information Across Network Layers

Cross-layer fusion combines low-level spatial information with high-level semantics to support small-object detection. Approaches include concatenation, addition, attention-based weighting, and feature fusion networks.

4-3-3- Fusion of Information Across Network Layers

  • Cross-layer fusion helps detect small objects that may be represented only in early layers by combining spatial detail with semantic information.Different layers capture edges, textures, and high-level semantics (Zhang, Zhang, et al., 2024; Z. Zhou & Zhu, 2024).Source: 4-3-3- Fusion of Information Across… · p. 21
  • Attention mechanisms dynamically weight features across layers to emphasize relevant information (Jiang et al., 2024; Zhang, Zhang, et al., 2024; F. Zhao et al., 2024).Spatial Attention Modules (SAM) emphasize relevant spatial regions, while Channel Attention Modules (CAM) emphasize relevant channels. Concatenation and addition are alternative fusion strategies.Source: 4-3-3- Fusion of Information Across… · p. 21
  • Feature Fusion Networks (FFNs) combine layer outputs into holistic representations, improving detection when small objects are poorly represented in individual layers.These networks combine high-level semantic information with low-level spatial information.Source: 4-3-3- Fusion of Information Across… · p. 21

4-4- Data Augmentation and Synthetic Data · 4-4-1- Data Augmentation Techniques for SOD

Data augmentation expands SOD training datasets through image transformations that simulate real-world conditions and emphasize small-scale features (Q. Liu et al., 2024).

4-4-1- Data Augmentation Techniques for SOD

4-4-2- Synthetic Data Generation for SOD … 4-5-3- Integration of Multi-Task and Transfer Learning

Synthetic data and augmentation address limited SOD training data, while multi-task and transfer learning share or reuse representations to improve detection. These approaches face synthetic–real domain gaps, conflicting task objectives, domain shift, and overfitting.

4-4-2- Synthetic Data Generation for SOD

  • Synthetic data generation creates new SOD training samples and pixel-perfect annotations, reducing collection and labeling costs (Q. Liu et al., 2024).GANs generate rare or diverse objects and backgrounds (D. Chen et al., 2024; Jing, Zhang, Li, et al., 2024b; Tian et al., 2024), while CGI controls object positioning, size, and texture (Zhang, Zhang, et al., 2024). Simulators and 3D rendering vary lighting, camera angles, and occlusions to mimic real-world conditions.Source: 4-4-2- Synthetic Data Generation for SOD · p. 22

4-4-3- Advantages of Data Augmentation and Synthetic Data in SOD

  • Augmentation and synthetic data improve generalization and robustness, alleviate data scarcity and class imbalance, and reduce dataset costs.They support unseen environments (Q. Liu et al., 2024), robustness to scale, occlusion, and lighting variations, and economical dataset expansion where collection is limited (Q. Liu et al., 2024; F. Zhao et al., 2024). Targeted oversampling increases representation of underrepresented small-object classes (F. Zhao et al., 2024).Source: 4-4-3- Advantages of Data Augmentation… · p. 23

4-4-4- Challenges and Research Directions

  • Synthetic–real visual gaps, excessive augmentation, and resource-intensive dataset generation limit synthetic-data approaches, motivating adaptation, careful tuning, and more realistic generation.Domain gaps impair real-world performance, and diverse synthetic datasets require substantial computation and expertise (S. Chen et al., 2024); unrealistic augmentation can harm training (S. Chen et al., 2024; Liang et al., 2020). Proposed directions include automated augmentation pipelines such as AutoAugment, diffusion-based realism improvements, and multimodal synthetic data for contextual learning.Source: 4-4-4- Challenges and Research… · p. 23

4-5-1- Multi-Task Learning for SOD

  • Multi-task learning shares features across related tasks to enrich small-object representations, mitigate overfitting, and reduce the computation required by separate models.Shared detection, segmentation, and classification backbones (M. Li et al., 2024), auxiliary SR, edge detection, and depth estimation (L. Zhou et al., 2024), multi-scale tasks, and task-specific heads provide complementary information. Conflicting objectives require balancing, such as dynamic loss weighting (S. Li et al., 2024), while diverse tasks increase architecture and training complexity.Source: 4-5-1- Multi-Task Learning for SOD · pp. 23–24

4-5-2- Transfer Learning for SOD

  • Transfer strategies include feature reuse, selective fine-tuning, domain-specific pretraining, knowledge distillation, cross-domain adaptation, and few-shot learning.ImageNet or COCO features transfer edges and textures (Butler & Leung, 2024; Y. Li, Yang, et al., 2024); typical fine-tuning freezes lower layers and adapts higher layers (Tian et al., 2024; W. Zhou et al., 2024), while aerial and medical pretraining improves domain-specific transferability (Jing, Zhang, Li, et al., 2024a; Zou et al., 2024). Knowledge distillation transfers teacher knowledge to lightweight students under resource constraints (Jobaer et al., 2025; W. Zhou et al., 2024), domain-invariant features support large-to-small-object transfer, and few-shot learning addresses very limited labels (B. Wang et al., 2024).Source: 4-5-2- Transfer Learning for SOD · pp. 24–25
  • Transfer learning reduces labeled-data requirements, accelerates convergence, and improves performance over training from scratch (Aibibu et al., 2023b; Butler & Leung, 2024).Knowledge from large-scale datasets also improves generalization to unseen scenarios. Source–target domain shift requires domain adaptation (Nabavi et al., 2023; Nabavi, et al., 2024), while fine-tuning on small, insufficiently varied target datasets risks overfitting.Source: 4-5-2- Transfer Learning for SOD · p. 25

4-5-3- Integration of Multi-Task and Transfer Learning

  • Combining multi-task and transfer learning fine-tunes pretrained multi-task models for SOD, with applications in aerial imagery, medical imaging, and crowded-scene surveillance.Applications include detecting vehicles, animals, and buildings in satellite or drone imagery, alongside lesions, tumors, and abnormalities in medical scans (C. Chen et al., 2024; W. Zhou et al., 2024). Surveillance applications include weapons and unattended bags in crowds (Cao et al., 2024; Z. Chen, Shi, et al., 2024; D. Liao et al., 2025b).Source: 4-5-3- Integration of Multi-Task and… · pp. 25–26
  • Future integration research targets automated task balancing, stronger domain adaptation, few-shot multi-task learning, and continual transfer learning.Dynamic task balancing addresses conflicting objectives (S. Li et al., 2024), domain adaptation addresses domain shifts, and few-shot integration targets SOD with limited examples. Continual transfer learning aims to accommodate new tasks and domains while retaining previously learned knowledge.Source: 4-5-3- Integration of Multi-Task and… · p. 26

5- Datasets and Evaluation Metrics · 5-1- Datasets

The section reviews small-object detection datasets for surveillance, autonomous driving, and aerial imagery, highlighting challenges from size variation, occlusion, and complex backgrounds.

5-1- Datasets

  • The surveyed datasets support real-world detection tasks, including surveillance, autonomous driving, and tracking, while addressing size variation, occlusion, and complex backgrounds.WSODD targets small or floating objects on water surfaces; KITTI provides 3D object annotations for autonomous-driving research.Source: 5-1- Datasets · pp. 26–27
  • 80 object categories and more than 2.5 million object instances are reported for one summarized dataset, illustrating the breadth of annotated detection data.Listed categories include people, animals, and vehicles; the supplied fragments do not identify this dataset.Source: 5-1- Datasets · p. 27
  • SODA includes small-object annotations for vehicles and pedestrians, with size categories such as extremely small and relatively small.Its stated focus includes improving detection in UAV and aerial imagery.Source: 5-1- Datasets · p. 27
  • BDD100K presents weather and lighting variations, occlusion, moving objects, and environmental complexity across urban streets and rural roads.Weather conditions include rain, snow, and sunshine.Source: 5-1- Datasets · p. 27
  • 100,000 images constitute BDD100K, which supports autonomous-driving research and object detection in real-world driving scenarios.The images are captured using mounted cameras.Source: 5-1- Datasets · p. 27
  • AI-TOD challenges include objects occupying very few pixels, occlusion, scale variation, low resolution, and complex backgrounds.Tiny objects can also be difficult to distinguish from noise, further complicating accurate detection.Source: 5-1- Datasets · p. 27

5-1-1- Aerial Image Datasets … 5-1-1-4- VEDAI

Aerial imagery datasets support small object detection in settings with variable object sizes, occlusion, and dense clustering. The reviewed datasets cover diverse detection conditions, remote-sensing robustness, rotated-box localization, and small-vehicle detection.

5-1-1- Aerial Image Datasets

  • Aerial imagery datasets are extensively used for small object detection because of their real-world relevance and challenges involving variable object sizes, occlusion, and dense clustering.Source: 5-1-1- Aerial Image Datasets · p. 28

5-1-1-1- VisDrone

  • 10 object categories in VisDrone (P. Zhu et al., 2018) span varying spatial densities and occlusion, with additional weather, lighting, and background challenges.Categories include pedestrians, bicycles, vehicles, and buses; backgrounds include urban and rural environments.Source: 5-1-1-1- VisDrone · p. 28
  • Over 10,209 images comprise VisDrone (P. Zhu et al., 2018), a widely adopted aerial small-object detection benchmark.The dataset contains training (6,471), validation (548), and testing (3,190) subsets.Source: 5-1-1-1- VisDrone · p. 28

5-1-1-2- DIOR

  • 23,463 images and 192,472 annotated instances across 20 object categories constitute DIOR (K. Li et al., 2020), a remote-sensing detection dataset.Images from multiple sensors under different conditions support testing detection robustness.Source: 5-1-1-2- DIOR · p. 28

5-1-1-3- DOTA

  • Over 15 object categories in DOTA, including small vehicles, ships, and storage tanks, use rotated bounding boxes for precise localization.DOTA-v1.0, v1.5, and v2.0 incrementally improve sample size and annotations.Source: 5-1-1-3- DOTA · p. 28

5-1-1-4- VEDAI

  • VEDAI targets multi-class small-vehicle detection in aerial imagery across varying resolutions and cluttered environments.Source: 5-1-1-4- VEDAI · p. 28

5-1-2- UAV-Based Datasets … 5-2-1- Metrics Commonly Used Across General Object Detection

The reviewed datasets span UAV, autonomous-driving, specialized, and synthetic settings, with low resolution, occlusion, and variable conditions recurring as challenges. Future directions emphasize richer datasets and annotations, while evaluation uses AP and mAP.

5-1-2- UAV-Based Datasets; 5-1-2-1- UAVDT

  • Over 100 videos in UAVDT (H. Yu et al., 2020) annotate vehicles across scenarios, with motion blur, small object sizes, and varying weather challenging detection.UAVs capture images from different altitudes and angles; UAVDT focuses on moving-object detection in settings including highways and parking lots.Source: 5-1-2-1- UAVDT · p. 29, 5-1-2- UAV-Based Datasets · p. 29

5-1-2-2- SODA-D; 5-1-2-3- DroneCrowd

  • SODA-D (C. Chen et al., 2017) covers UAV small objects across weather and lighting conditions; DroneCrowd targets dense, occluded human crowds.SODA-D annotates nine classes, including pedestrians, cars, and bicycles, using large, medium, small, and very small size categories. DroneCrowd supports detecting and tracking small humans in crowded urban areas.Source: 5-1-2-2- SODA-D · p. 29, 5-1-2-3- DroneCrowd · p. 29

5-1-3- Autonomous Driving Datasets; 5-1-3-1- BDD100K; 5-1-3-2- KITTI

  • 100,000 diverse driving scenes in BDD100K (F. Yu et al., 2018) support multi-class detection; KITTI (Y. Liao et al., 2021) emphasizes 3D detection.BDD100K covers traffic lights, pedestrians, and cars under varying weather, lighting, and urban/rural settings. KITTI includes small-object categories such as pedestrians and bicycles.Source: 5-1-3-1- BDD100K · p. 29, 5-1-3-2- KITTI · pp. 29–30

5-1-4- Specialized Datasets for SOD; 5-1-4-1- TinyPerson; 5-1-4-2- WSODD; 5-1-4-3- AI-TOD

  • TinyPerson (X. Yu et al., 2020), WSODD (Z. Zhou et al., 2021), and AI-TOD (J. Wang et al., 2020) target humans, maritime objects, and aerial objects.TinyPerson provides high-resolution annotations for small pedestrians for surveillance and crowd analysis; WSODD includes small boats and buoys affected by waves and environmental occlusion. AI-TOD annotates very small objects in real-world aerial imagery.Source: 5-1-4-1- TinyPerson · p. 30, 5-1-4-2- WSODD · p. 30, 5-1-4-3- AI-TOD · p. 30

5-1-5- Synthetic Datasets; 5-1-5-1- Motion Blur Synthetic Dataset; 5-1-5-2- Self-Collected Datasets

5-1-6- Comparative Insights and Challenges

  • Low resolution, occlusion, clutter, environmental variability, and application-specific requirements constrain small-object datasets.Few-pixel objects impede feature extraction (Zhang, Ye, et al., 2024), while dense arrangements hinder detection (H. Wang et al., 2024). Weather, lighting, and orientation add complexity, and aerial, driving, and industrial applications require tailored datasets.Source: 5-1-6- Comparative Insights and… · p. 31

5-1-7- Future Directions

  • Future dataset priorities include underrepresented domains, larger and more diverse collections, high-fidelity synthesis, multimodal sensing, and more precise annotations.Thermal imaging and LiDAR complement RGB to enhance detection accuracy (Z. Chen, Shi, et al., 2024; Xu et al., 2024). Polygon masks and 3D bounding boxes can improve labeling precision and detection in complex environments.Source: 5-1-7- Future Directions · p. 31

5-2- Evaluation Metrics; 5-2-1- Metrics Commonly Used Across General Object Detection

  • SOD evaluation requires metrics accounting for small-object challenges; AP measures per-class precision–recall curve area, while mAP averages AP across classes.Reported metrics include mAP(50), mAP(75), and mAP(50:95); AP(50:95) averages AP across IoUs from 0.50 to 0.95 in 0.05 steps. mAP is the primary multi-class comparison metric and accommodates datasets with varying object sizes.Source: 5-2-1- Metrics Commonly Used Across… · pp. 31–32, 5-2- Evaluation Metrics · p. 31

5-2-2- Metrics for Size-Specific Detection … 8- Conclusion

The survey combines scale-specific, tracking, and efficiency metrics with applications spanning aerial monitoring, autonomous systems, industry, healthcare, agriculture, and maritime environments. Future priorities include efficient deployment, stronger feature representations, domain generalization, and model interpretability.

5-2-2- Metrics for Size-Specific Detection; 5-2-3- Tracking-Specific Metrics

  • Size-specific AP distinguishes object scales, while tracking metrics assess small-object continuity, localization at low IoU thresholds, and temporal consistency.COCO-style APS_SS, APM_MM, and APL_LL evaluate small, medium, and large objects (C. Chen et al., 2024); APT_TT evaluates objects smaller than COCO’s small category in datasets such as TinyPerson (Jing, Zhang, Li, et al., 2024b; K. Liu et al., 2025). T-AP10, T-AP15, and T-AP20 use IoU thresholds of 10%, 15%, and 20%, respectively, while T-mAP accounts for consistency across frames.Source: 5-2-2- Metrics for Size-Specific… · p. 32, 5-2-3- Tracking-Specific Metrics · p. 32

5-2-4- Core Metrics for Precision and Recall; 5-2-5- Specialized Metrics for Small Object Datasets; 5-2-7- Computational Efficiency Metrics

  • Precision, Recall, and F1-Score assess detection quality, while PPN, SNR, CV, FPS, and inference time characterize processing efficiency or challenging-image performance.Precision addresses false positives, Recall captures relevant instances, and F1-Score balances them (Q. Liu et al., 2024); PPN measures processed-pixel proportion, while SNR and CV assess detection quality under noise or low resolution. FPS accompanies AP or mAP to expose accuracy–speed trade-offs (K. Liu et al., 2025), and inference time measures average image or frame processing time for resource-constrained applications.Source: 5-2-4- Core Metrics for Precision and… · p. 32, 5-2-5- Specialized Metrics for Small… · p. 33, 5-2-7- Computational Efficiency Metrics · p. 33

5-2-6- Dataset-Specific Metric Implementation

  • Dataset-specific evaluation combines general AP measures with size-specific scores and, where needed, custom recall or image-quality metrics.COCO and PASCAL VOC use AP, AP(50), AP(75), and size-specific scores; VisDrone and TinyPerson emphasize APT_TT, mAP, and APS_SS (Z. Chen, Ji, et al., 2024; Jing, Zhang, Li, et al., 2024b; K. Liu et al., 2025). DIOR and DOTA use mAP, AP(50), AP(75), and APS_SS for complex remote-sensing scenes (Gao, Wang, et al., 2024; Shi et al., 2024; Z. Zhou & Zhu, 2024), while specialized datasets such as SODA-D or WSODD may use mAR and image-quality metrics.Source: 5-2-6- Dataset-Specific Metric… · p. 33

6- Applications and Real-world Use Cases; 6-1- Remote Sensing and Aerial Surveillance; 6-2- UAV and Drone Applications; 6-9- Intelligent Transportation Systems

  • Aerial and transportation applications use SOD for urban monitoring, disaster response, surveillance, traffic analysis, and infrastructure inspection.Satellite and aerial imagery support urban planning, damage assessment, wildlife and environmental monitoring, reconnaissance, and maritime surveillance; UAV applications additionally include fire warnings, search and rescue, and crop monitoring. Transportation uses include weather- and lighting-dependent traffic-sign recognition (Jiang et al., 2024; H. Wang et al., 2024), pedestrian-density monitoring (D. Liao et al., 2025b), and detecting faults or obstacles on railways and roads.Source: 6-1- Remote Sensing and Aerial… · p. 34, 6-2- UAV and Drone Applications · p. 34, 6- Applications and Real-world Use Cases · p. 34

6-3- Autonomous Systems; 6-4- Industrial Applications; 6-5- Surveillance and Security; 6-7- Medical Imaging and Biological Analysis

  • Autonomous, industrial, security, and medical systems use SOD to identify navigation hazards, manufacturing defects, suspicious activity, and minute biological abnormalities.Autonomous vehicles detect distant signs, pedestrians, bicycles, and obstacles (J. Liu et al., 2024); USVs and drones use detection for navigation (B. Wang et al., 2024), while SLAM benefits mapping accuracy. Other applications include component-defect detection, aerospace and cable inspection, CCTV and border monitoring (F. Zhao et al., 2024), and detecting small tumors or strokes alongside tissue and cellular analysis.Source: 6-3- Autonomous Systems · p. 34, 6-4- Industrial Applications · p. 35, 6-5- Surveillance and Security · p. 35

6-6- Environmental and Natural Resource Monitoring; 6-8- Maritime and Oceanographic Applications; 6-10- Specialized Applications in Agriculture and Forestry

  • Agricultural, environmental, and maritime applications monitor small pests, resource changes, forest conditions, navigation obstacles, and coastal threats.Agriculture targets pests, weeds, and crop diseases, while forestry tracks health, wildlife, and illegal logging; resource surveys monitor mineral deposits, water levels, and subtle geological features (F. Feng et al., 2024). Maritime applications detect buoys, debris, and vessels for navigation, track marine life and oceanographic changes, and monitor coastal or offshore threats.Source: 6-6- Environmental and Natural… · p. 35, 6-8- Maritime and Oceanographic… · pp. 35–36, 6-10- Specialized Applications in… · p. 36

7- Future Directions

  • Future priorities combine efficient architectures, stronger features, synthetic data, domain adaptation, improved training objectives, and temporal modeling for robust deployment.Lightweight backbones, parameter sharing, multiscale fusion, dynamic feature focusing, SR, and CNN–transformer hybrids target efficiency and feature quality; GANs and diffusion models generate varied training data, while domain adaptation targets unseen conditions. Improved label assignment and losses address imbalance, localization, and false detections; temporal context targets fast-moving objects, emerging applications include SLAM, and pruning, quantization, and architectural refinements support resource-constrained deployment.Source: 7- Future Directions · pp. 36–37

8- Conclusion

  • Small-object detection advances through multiscale features, attention, super-resolution, lightweight architectures, transformers, and knowledge distillation, but efficiency, dataset generalization, and interpretability remain unresolved.Limited resolution, background interference, and class imbalance motivate these approaches, while tailored datasets and evaluation metrics support benchmarking. Multimodal data and domain-specific strategies broaden applications.Source: 8- Conclusion · p. 38

Related papers

Loading 2503.20516v1…