Abstract:Recent years have witnessed rapid developments on computer vision, however, there are still challenges in detecting tiny objects in a large-scale background. The tiny objects knowledge become sparse and weak due to their tiny size, which makes the tiny objects difficult to be detected with the common approaches. In this paper, a new network named Specific Characteristics based Feature Rescaling and Fusion (SFRF) is designed to detect tiny persons in a broad horizon and massive background. Different from the methods in general, a Nonparametric Adaptive Dense Perceiving Algorithm (NADPA) is designed to automatically select and generate a new resized feature map with the high density distribution of tiny objects. Then, a method called Many-For-One strategy is used for feature fusion of the feature pyramid network (FPN) layers to improve the feature representation and detection. Finally, an ensemble model method named hierarchical Coarse-to-fine mechanism is designed based on the proposed methods to further improve the performance. The experiments demonstrate that the proposed approach achieves an obvious performance improvement on tiny object detection than the existing approaches, and our approach has been awarded as the 1st-place in the first large-scale Tiny Object Detection (TOD) challenge.

What problem does this paper attempt to address?

### Problems Addressed by the Paper This paper aims to address the problem of small object detection in large-scale contexts. Specifically, small objects perform poorly in common detection methods due to their extremely small size, as these methods cannot effectively capture and utilize the feature information of small objects. The knowledge of small objects becomes sparse and weakened, making them difficult to detect. ### Solution Overview To tackle this challenge, the authors designed a new network structure called Specific Characteristics based Feature Rescaling and Fusion (SFRF). The SFRF network mainly improves the detection performance of small objects through the following three aspects: 1. **Nonparametric Adaptive Dense Perceiving Algorithm (NADPA)**: - A nonparametric adaptive dense perceiving algorithm is designed to automatically select and generate new rescaled feature maps of densely distributed small objects. - This algorithm can effectively identify dense areas in the image, thereby improving the detection accuracy of small objects. 2. **Many-For-One Strategy**: - Used for feature fusion in the Feature Pyramid Network (FPN) layers to improve feature representation and detection performance. - By fusing multiple high-level features into low-level features, the semantic information of low-level features is enhanced, thereby improving the detection effect of small objects. 3. **Hierarchical Coarse-to-Fine Mechanism**: - An integrated model method based on the above methods is designed to further enhance overall performance. - Through a multi-level coarse-to-fine mechanism, the detection results are gradually optimized, improving the stability and accuracy of the model. ### Experimental Results Experiments show that the proposed SFRF method achieves significant performance improvement in small object detection tasks, with an average precision (mAP) increase of 4.9% compared to existing methods, and a 14.33% improvement over the baseline model. Additionally, this method won first place in the large-scale small object detection (TOD) challenge. ### Main Contributions 1. **Proposed a new small object detection method SFRF**, effectively solving the problem of small object detection in large-scale contexts. 2. **Designed key technical components**, including NADPA, Many-For-One Strategy, and Hierarchical Coarse-to-Fine Mechanism, significantly enhancing detection performance. 3. **Performed excellently in practical applications**, achieving outstanding results in competitions and validating its effectiveness and efficiency on multiple datasets. ### Conclusion Small object detection is a challenging research field. The SFRF method proposed in this paper significantly improves the detection performance of small objects through innovative feature rescaling and fusion techniques. Future research will further extend to other types of small object detection tasks.

Feature Rescaling and Fusion for Tiny Object Detection

Effective Fusion Factor in FPN for Tiny Object Detection

Extended Feature Pyramid Network for Small Object Detection

Small Object Detection using Multi-scale Feature Fusion and Attention

A Small-Sized Object Detection Oriented Multi-Scale Feature Fusion Approach With Application to Defect Detection

FFEDet: Fine-Grained Feature Enhancement for Small Object Detection

ALFPN: Adaptive Learning Feature Pyramid Network for Small Object Detection

Dynamic Feature Focusing Network for Small Object Detection

Tiny object detection with context enhancement and feature purification

Small Object Detection Method Based on Adaptive Spatial Parallel Convolution and Fast Multi-Scale Fusion

A Small Target Detection Method Based on Feature Enhancement and Positioning Optimization

CA2Det: Cascaded Adaptive Fusion Pyramid Network Based on Attention Mechanism for Small Object Detection

HS-FPN: High Frequency and Spatial Perception FPN for Tiny Object Detection

Regression-Guided Refocusing Learning With Feature Alignment for Remote Sensing Tiny Object Detection

Feature Super-Resolution Fusion With Cross-Scale Distillation for Small-Object Detection in Optical Remote Sensing Images

Attentional feature pyramid network for small object detection

Multi-Scale Object Detection Using Feature Fusion Recalibration Network

SSPNet: Scale Selection Pyramid Network for Tiny Person Detection From UAV Images

Feature Implicit Enhancement Via Super-Resolution for Small Object Detection.

Enhanced semantic feature pyramid network for small object detection

A DeNoising FPN With Transformer R-CNN for Tiny Object Detection