Small object detection: A comprehensive survey on challenges, techniques and real-world applications

This paper, “Small object detection: A comprehensive survey on challenges, techniques and real-world applications” (Intelligent Systems with Applications, 2025), addresses a broad but important research question: what are the dominant challenges in small object detection (SOD), what deep-learning techniques have been proposed to overcome them, how are SOD systems evaluated (datasets and metrics), and where do these methods work in practice. The question matters because SOD is a prerequisite for many safety- and mission-critical systems—e.g., autonomous driving (distant pedestrians, cyclists), surveillance (unattended objects in clutter), medical imaging (tiny lesions), and remote sensing (small vehicles/structures). In these settings, small objects often determine whether a system can act correctly; missing them can lead to catastrophic outcomes.

The survey’s significance lies in its attempt to consolidate the state of the art specifically for the most recent period (articles published in Q1 journals during 2024–2025 indexed in Scopus with “Small Object Detection” in the title). It frames SOD as distinct from general object detection: small objects contain limited spatial and contextual information, are easily confused with background clutter, and are frequently underrepresented in training data. The authors emphasize that progress in SOD is not only about model architectures, but also about data (augmentation/synthetic generation), evaluation protocols (size-specific metrics), and deployment constraints (lightweight and real-time performance).

Methodologically, this is a narrative/technical survey rather than an empirical study. The “study design” is therefore a literature selection and synthesis process. The authors define SOD and categorize challenges, then review deep-learning techniques and trends, and finally summarize commonly used datasets and evaluation metrics. The paper does not report a sample size in the statistical sense (no experimental cohort), but it does provide a quantitative bibliometric-style result: in their analysis of recent trends, optimized backbone architectures account for 23.1% of the reviewed literature, attention mechanisms for 18.5%, feature extraction enhancement for 16.9%, feature fusion optimization for 15.4%, advanced learning strategies for 13.8%, and multi-scale hybrid attention/high-resolution feature alignment for 12.3%. This gives a sense of where the research community’s effort is concentrated.

A central part of the survey is the decomposition of SOD challenges. The authors highlight (i) limited appearance information and occlusion, (ii) localization difficulty and scale variation (including poor IoU with anchors and receptive-field misalignment), (iii) inefficiency in feature learning and background interference (downsampling reduces discriminative details; low signal-to-noise ratio), (iv) limitations of popular detectors originally designed for larger objects (loss of fine-grained detail; computational cost; multi-task imbalance), (v) high computational costs and hardware constraints for high-resolution imagery and edge deployment, and (vi) inconsistent performance across scales and datasets (anchor/grid mismatch, feature disappearance, and dense small-object scenes). They also connect these challenges to practical deployment settings such as UAV/edge computing and industrial environments (e.g., coal mines with dust/low illumination).

On techniques, the survey organizes recent deep-learning approaches into several families. First are architecture-level improvements: optimized backbones and lightweight designs (e.g., FFEDet, KDSMALL), feature extraction enhancement (multi-resolution modules, dilated convolutions, large-kernel spatial pyramid pooling), and feature fusion/multi-scale detection (FPN variants, bidirectional feature pyramids, attention-guided fusion). Second are attention mechanisms, including channel/spatial attention and transformer-based attention, used to suppress irrelevant background and emphasize small-object cues. Third are training and learning strategies: knowledge distillation (KD) to improve efficiency, self-supervised learning to reduce reliance on labels, reinforcement learning in some contexts, and advanced data augmentation and synthetic data generation (GANs, diffusion models, simulators/3D rendering). Fourth are problem-specific solutions: anchor-free detectors, two-stage designs with low-resolution localization and high-resolution classification, and specialized losses/label assignment to handle imbalance and improve localization (e.g., size-specific IoU losses, balanced label assignment).

The survey also provides concrete examples of representative methods (summarized in a table). While the paper does not provide per-method numerical performance comparisons (e.g., mAP gains with p-values), it does describe the key technical components of many systems. Examples include MICPL (motion-inspired cross-pattern learning for satellite videos), ScorePillar (real-time small object detection using pillar scoring of LiDAR), LA-YOLO (bidirectional adaptive feature fusion for insulator defects), KDSMALL (EfficientNet backbone plus CBAM attention and KD), and various YOLOv8-based lightweight SOD variants for UAV/autonomous driving/remote sensing. These examples support the survey’s broader claim that the dominant engineering pattern is: preserve high-resolution/localization information while enriching semantics via multi-scale fusion and attention, and then address data scarcity/imbalance via augmentation, synthetic data, and distillation.

For evaluation, the authors review datasets and metrics. They emphasize that standard object detection metrics must be adapted for SOD. Common metrics include AP and mAP with COCO-style IoU thresholds (e.g., AP(50), AP(75), and AP(50:95)). For small objects, they discuss size-specific AP metrics such as APS_SS, APM_MM, and APL_LL (small/medium/large bins as defined by pixel area thresholds in COCO-like protocols), and tiny-object metrics such as APT_TT for datasets like TinyPerson. For tracking-oriented SOD, they mention T-AP10/T-AP15/T-AP20 and T-mAP to account for temporal consistency. They also note computational efficiency metrics such as FPS and inference time, and application-specific measures like processed pixel number percentage (PPN) and noise-related statistics (SNR, coefficient of variation).

The dataset review is extensive and covers both general detection benchmarks and SOD-focused resources. The paper lists major aerial/remote-sensing datasets (VisDrone, DIOR, DOTA, VEDAI), UAV video datasets (UAVDT), crowd-focused aerial datasets (DroneCrowd), and tiny-object/person datasets (TinyPerson, SODA variants). It also includes autonomous driving datasets (BDD100K, KITTI) and specialized domains (water surface object detection WSODD; tiny object detection AI-TOD). The authors provide dataset sizes and annotation counts for several key datasets—for example, VisDrone has 10,209 images total with 6,471 train, 548 validation, and 3,190 test; DOTA has 2,806 high-resolution aerial images with 188,282 object instances; and TinyPerson has 72,000 images with annotations for small pedestrian detection.

Limitations are not presented as a formal “limitations” section with explicit methodological constraints, but limitations are apparent from the nature of the work: it is a survey, so it does not conduct controlled comparisons across methods, does not standardize evaluation across papers, and does not quantify effect sizes of technique families on a common benchmark. Additionally, the selection criterion (“small object detection” in the title and Q1 journals during 2024–2025) may exclude relevant SOD work that uses different terminology (e.g., “tiny object detection,” “small target detection”) or appears in non-Q1 venues.

Practical implications are a major focus. The authors argue that SOD progress should be judged not only by accuracy but also by robustness to domain shift (weather, lighting, blur), and by real-time feasibility on resource-constrained platforms (UAVs, edge devices, FPGA). They highlight who should care: researchers choosing architectures and training strategies; practitioners in surveillance, remote sensing, medical imaging, industrial inspection, and agriculture who need reliable detection of small cues; and system designers who must balance accuracy with latency and compute.

Overall, the paper’s core contribution is a structured synthesis of SOD challenges, the deep-learning techniques most commonly used to address them (multi-scale fusion, attention, SR, KD, augmentation/synthetic data, transformer and lightweight architectures), and the evaluation ecosystem (datasets and size-specific metrics). It closes by pointing to open problems: robust domain adaptation, better feature fusion strategies, and real-time optimization, especially for deployment in dynamic and noisy environments.