SARDet-100K: Towards Open-Source Benchmark and ToolKit for Large-Scale SAR Object Detection (2403.06534v2)

Abstract: Synthetic Aperture Radar (SAR) object detection has gained significant attention recently due to its irreplaceable all-weather imaging capabilities. However, this research field suffers from both limited public datasets (mostly comprising <2K images with only mono-category objects) and inaccessible source code. To tackle these challenges, we establish a new benchmark dataset and an open-source method for large-scale SAR object detection. Our dataset, SARDet-100K, is a result of intense surveying, collecting, and standardizing 10 existing SAR detection datasets, providing a large-scale and diverse dataset for research purposes. To the best of our knowledge, SARDet-100K is the first COCO-level large-scale multi-class SAR object detection dataset ever created. With this high-quality dataset, we conducted comprehensive experiments and uncovered a crucial challenge in SAR object detection: the substantial disparities between the pretraining on RGB datasets and finetuning on SAR datasets in terms of both data domain and model structure. To bridge these gaps, we propose a novel Multi-Stage with Filter Augmentation (MSFA) pretraining framework that tackles the problems from the perspective of data input, domain transition, and model migration. The proposed MSFA method significantly enhances the performance of SAR object detection models while demonstrating exceptional generalizability and flexibility across diverse models. This work aims to pave the way for further advancements in SAR object detection. The dataset and code is available at https://github.com/zcablii/SARDet_100K.

• The paper presents SARDet-100K, a unified COCO-level dataset combining 10 SAR datasets with nearly 250K annotations across six categories.
• It introduces the MSFA pretraining framework that integrates handcrafted features and multi-stage training to bridge domain gaps from RGB to SAR data.
• Experimental evaluations on benchmarks like SSDD and HRSID show significant performance gains over traditional ImageNet pretraining methods.

SARDet-100K: An Open-Source Benchmark for Large-Scale SAR Object Detection

The paper "SARDet-100K: Towards Open-Source Benchmark and Toolkit for Large-Scale SAR Object Detection" tackles critical challenges in SAR (Synthetic Aperture Radar) object detection by introducing a large-scale dataset and a novel pretraining framework. The scarcity of large and diverse public SAR datasets has hindered the advancement of object detection in SAR imagery, which is uniquely valuable due to its robust all-weather capabilities. The SARDet-100K dataset and the Multi-Stage with Filter Augmentation (MSFA) pretraining framework aim to bridge these gaps.

Key Contributions

1. SARDet-100K Dataset: This is the first COCO-level dataset for SAR object detection, amalgamating 10 existing datasets into a unified, standardized format. It encompasses 116,598 images annotated with 245,653 instances across six categories: Ship, Tank, Bridge, Harbour, Aircraft, and Car. This comprehensive dataset addresses the limitations of previous datasets that were often small in scale and single-category focused. SARDet-100K facilitates rigorous evaluation and development of SAR object detection models by providing a diverse and large-scale benchmark.
2. MSFA Pretraining Framework: A significant contribution of the paper is its solution to the domain and model gaps encountered when applying models pretrained on natural RGB datasets to SAR data. The MSFA framework comprises Filter Augmented Input and Multi-Stage Pretraining strategies. By incorporating handcrafted feature descriptors into the input data, the framework reduces visual domain discrepancies, enhancing cross-domain transferability. Additionally, the multi-stage approach includes a second pretraining on a large-scale optical remote sensing dataset, further aligning the model structure with SAR data requirements.

Experimental Evaluation

The experiments demonstrate notable improvements in detection performance using the proposed MSFA framework. Compared to traditional ImageNet pretraining methods, the MSFA approach achieves superior results across various detection models and backbones. Specifically, the framework shows strong performance on benchmarks like SSDD and HRSID, surpassing state-of-the-art methods with significant margins.

Implications and Future Directions

The introduction of SARDet-100K and MSFA marks a significant advancement in SAR object detection research. The dataset provides a robust foundation for model training and benchmarking, enabling more accurate and generalizable models. The MSFA framework's adaptability across different architectures highlights its potential applicability beyond SAR to other domains with similar challenges in domain adaptation.

Future research could explore further enhancements to the pretraining strategies, such as incorporating additional domain-specific features or optimizing the model architecture for SAR data. Additionally, ongoing development and expansion of datasets like SARDet-100K could facilitate more nuanced advancements in this field.

In conclusion, the paper's contributions address two core challenges in SAR object detection: the lack of a publicly accessible, large-scale benchmark and the transferability gap from RGB to SAR datasets. By providing both a standardized dataset and an innovative pretraining approach, the authors set a new standard for research and development in SAR object detection.