SARDet-100K: Towards Open-Source Benchmark and Toolkit for Large-Scale SAR Object Detection This paper introduces SARDet-100K, a large-scale, multi-class SAR object detection dataset and a novel pretraining framework called Multi-Stage with Filter Augmentation (MSFA). The dataset is created by merging 10 existing SAR detection datasets, resulting in a comprehensive collection of 116,598 images and 245,653 object instances across six categories: Ship, Aircraft, Car, Bridge, Tank, and Harbour. SARDet-100K is the first COCO-level dataset for SAR object detection, providing a rich resource for research and development. The MSFA framework addresses the domain and model gaps between pretraining on RGB datasets and finetuning on SAR datasets. It incorporates filter augmented input, which uses handcrafted features to enhance the transferability of pretrained knowledge. The framework also includes a multi-stage pretraining strategy that leverages a large-scale optical remote sensing dataset to bridge the domain gap between natural images and SAR images. Experiments show that MSFA significantly improves SAR object detection performance, achieving state-of-the-art results on benchmark datasets. The framework is generalizable and applicable across various deep neural networks, demonstrating robust performance across different detection frameworks and backbones. The paper also highlights the challenges in SAR object detection, including limited public datasets and inaccessible source code. The proposed dataset and framework aim to address these issues, providing a valuable resource for further advancements in SAR object detection research. The dataset and code are publicly available at https://github.com/zcablii/SARDet_100K.SARDet-100K: Towards Open-Source Benchmark and Toolkit for Large-Scale SAR Object Detection This paper introduces SARDet-100K, a large-scale, multi-class SAR object detection dataset and a novel pretraining framework called Multi-Stage with Filter Augmentation (MSFA). The dataset is created by merging 10 existing SAR detection datasets, resulting in a comprehensive collection of 116,598 images and 245,653 object instances across six categories: Ship, Aircraft, Car, Bridge, Tank, and Harbour. SARDet-100K is the first COCO-level dataset for SAR object detection, providing a rich resource for research and development. The MSFA framework addresses the domain and model gaps between pretraining on RGB datasets and finetuning on SAR datasets. It incorporates filter augmented input, which uses handcrafted features to enhance the transferability of pretrained knowledge. The framework also includes a multi-stage pretraining strategy that leverages a large-scale optical remote sensing dataset to bridge the domain gap between natural images and SAR images. Experiments show that MSFA significantly improves SAR object detection performance, achieving state-of-the-art results on benchmark datasets. The framework is generalizable and applicable across various deep neural networks, demonstrating robust performance across different detection frameworks and backbones. The paper also highlights the challenges in SAR object detection, including limited public datasets and inaccessible source code. The proposed dataset and framework aim to address these issues, providing a valuable resource for further advancements in SAR object detection research. The dataset and code are publicly available at https://github.com/zcablii/SARDet_100K.