FCOSR: A Simple Anchor-free Rotated Detector for Aerial Object Detection (2111.10780v2)

Abstract: Existing anchor-base oriented object detection methods have achieved amazing results, but these methods require some manual preset boxes, which introduces additional hyperparameters and calculations. The existing anchor-free methods usually have complex architectures and are not easy to deploy. Our goal is to propose an algorithm which is simple and easy-to-deploy for aerial image detection. In this paper, we present a one-stage anchor-free rotated object detector (FCOSR) based on FCOS, which can be deployed on most platforms. The FCOSR has a simple architecture consisting of only convolution layers. Our work focuses on the label assignment strategy for the training phase. We use ellipse center sampling method to define a suitable sampling region for oriented bounding box (OBB). The fuzzy sample assignment strategy provides reasonable labels for overlapping objects. To solve the insufficient sampling problem, a multi-level sampling module is designed. These strategies allocate more appropriate labels to training samples. Our algorithm achieves 79.25, 75.41, and 90.15 mAP on DOTA1.0, DOTA1.5, and HRSC2016 datasets, respectively. FCOSR demonstrates superior performance to other methods in single-scale evaluation. We convert a lightweight FCOSR model to TensorRT format, which achieves 73.93 mAP on DOTA1.0 at a speed of 10.68 FPS on Jetson Xavier NX with single scale. The code is available at: https://github.com/lzh420202/FCOSR

An Overview of FCOSR: A Simple Anchor-free Rotated Detector for Aerial Object Detection

The paper presents FCOSR, a one-stage anchor-free rotated object detector designed to simplify and enhance aerial image detection. The FCOSR model builds upon the FCOS architecture to directly predict the center point, width, height, and angle of an oriented bounding box (OBB), thus addressing challenges associated with the arbitrary orientation and dense arrangement of objects frequently encountered in aerial imagery.

Core Methodology

FCOSR circumvents traditional anchor-based methodologies by eliminating the need for cumbersome preset boxes, thereby reducing hyperparameter dependencies and computational overhead. The approach introduces a refined label assignment strategy based on a 2D Gaussian distribution to improve sample quality during training. Several key innovations in label assignment techniques have been integrated:

• Ellipse Center Sampling (ECS): This method enhances the traditional center sampling strategy by using elliptical sampling regions, which better adapt to the orientation and shape of detected objects than conventional square areas.
• Fuzzy Sample Label Assignment (FLA): By incorporating a Gaussian distribution as a distance measure between sampling points and target centers, this strategy effectively resolves ambiguities in sample labeling caused by overlapping objects.
• Multi-Level Sampling (MLS): An auxiliary sampling method addresses insufficient sampling issues, particularly for objects with large aspect ratios, by supplementing the coverage in lower-level feature maps.

Overall, FCOSR's architecture remains streamlined, comprising only convolutional layers, making it flexible and suitable for deployment across various platforms.

Performance and Results

The FCOSR model demonstrates competitive performance, achieving significant mean Average Precision (mAP) figures: 79.25% on DOTA1.0, 75.41% on DOTA1.5, and 90.15% on HRSC2016 datasets. The evaluations reveal that FCOSR outperforms existing state-of-the-art methods in single-scale testing scenarios. Additionally, a lightweight TensorRT version of FCOSR achieves 73.93 mAP at 10.68 FPS on a Jetson Xavier NX. These results underscore FCOSR's superior detection efficiency and deployment versatility.

A series of ablation experiments confirm the efficacy of individual methodological improvements, particularly highlighting the performance gains from rotation augmentation, ECS, FLA, and MLS modules.

Implications and Future Directions

The paper contributes to the object detection field by enhancing the practicality and deployment efficiency of anchor-free methodologies. By demonstrating competitive performance without complex architectures, FCOSR offers an attractive alternative for applications requiring rapid processing and limited computational resources, such as real-time aerial surveillance and remote sensing.

Looking ahead, further research could explore optimizing the ECS and MLS strategies for diverse environmental settings and object classes, amplifying the model's adaptability and precision across broader datasets. Moreover, advancing FCOSR's framework could facilitate improved edge device compatibility, streamlining integrated applications of AI in geographic and spatial analytics.

The continued innovation in anchor-free detection models like FCOSR holds promise for advancing the detection capabilities in diverse AI-driven applications, ensuring that aerial imagery analysis remains both efficient and effective in the face of evolving technological landscapes.