PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yuxuan Li, Xiang Li, Weijie Li, Qibin Hou, Li Liu, Ming-Ming Cheng, Jian Yang

arXiv:2403.06534v3cs.CVcs.AIcs.CEcs.LG

TL;DR

SAR object detection lacks large, diverse public datasets and accessible code, and RGB pretraining differs from SAR finetuning in data domain and model structure. The paper builds SARDet-100K and proposes MSFA, a multi-stage pretraining framework. MSFA is reported to improve detection performance and generalize across models, with new state-of-the-art results on benchmark datasets.

  • Problem

    SAR detection research faces limited public datasets and inaccessible code, alongside substantial data-domain and model-structure gaps between RGB pretraining and SAR finetuning.

    Source: §1 · pp. 1–2

  • Method

    The paper standardizes ten existing SAR datasets into SARDet-100K and proposes MSFA, which addresses input data, domain transition, and model migration during pretraining.

    Source: §1 · p. 2

  • Results

    MSFA is reported to improve SAR detection and generalize across models, achieving mAP@50 of 97.9% on SSDD and 83.7% on HRSID.

    Source: §1 · p. 2, §5.4 · p. 9, §7 · p. 10

  • Takeaways & Limitations

    SARDet-100K provides a large, diverse benchmark for SAR object detection research, while MSFA offers a pretraining framework applicable across different models.

    Source: §1 · p. 2

    The paper’s scope is limited to supervised pretraining; it identifies semi-supervised, weakly supervised, and unsupervised domain transfer as future directions.

    Source: §6 · p. 10

Abstract

from arXiv · show

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.

1 Introduction

SAR object detection is growing, but limited public datasets and inaccessible code hinder research, while RGB-to-SAR pretraining and finetuning face domain and model gaps. The paper introduces SARDet-100K and MSFA to address these challenges, reporting effectiveness and generalizability across models.

1 Introduction

  • Existing SAR detection datasets are often small, single-category, and paired with inaccessible source code, limiting reproducibility and fair comparison.The paper also identifies a domain gap between natural RGB and SAR imagery and a model gap between pretrained backbones and downstream detection frameworks.Source: §1 · p. 2
  • SARDet-100K combines public datasets into a COCO-level, six-category resource, while MSFA addresses data-domain transition and model-migration gaps in pretraining.The dataset and code are released to support research and comparisons.Source: §1 · pp. 2–3
  • MSFA tackles pretraining-to-finetuning gaps through feature-space input transformation, an optical remote-sensing transition bridge, and whole-detector model migration.The handcrafted descriptors are described as noise-robust and as narrowing RGB–SAR input-distribution differences.Source: §1 · p. 2
  • MSFA is reported to reduce pretraining-to-finetuning domain and model gaps and to generalize across modern deep neural networks.The introduction characterizes the framework as effective and applicable across various models.Source: §1 · p. 2

2 Related Work

Prior work uses handcrafted features in SAR analysis and specialized detection architectures, but the paper identifies limited use of input-level adaptation and overlooked pretraining-to-SAR gaps. It positions its contribution as a pretraining strategy for those challenges.

2.1 SAR Images and Handcraft Features

  • Handcrafted descriptors such as HOG, Canny, GRE, Haar-like features, and WST have been used for SAR feature extraction, but their role in domain adaptation and modern object detection remains underexplored.The paper distinguishes its focus from simply injecting handcrafted features into neural networks.Source: §2.1 · p. 3

2.2 SAR Object Detection

  • SAR detectors commonly address small objects, speckle noise, and sparse information through multiscale fusion, attention, and context-oriented network designs.Examples include MGCAN, MSSDNet, and SEFEPNet for feature fusion, and PADN and EWFAN for attention.Source: §2.2 · p. 3
  • Most prior work emphasizes network-structure changes or ImageNet-pretrained backbones, leaving input-level noise handling and domain and model gaps insufficiently addressed.The gaps concern natural-scene pretraining versus SAR finetuning and backbone pretraining versus the full detection framework.Source: §2.2 · p. 3

3 A New Benchmark Dataset for SAR Object Detection

High-resolution SAR data can be costly, sensitive, and difficult to access, while small, homogeneous datasets can bias evaluation. SARDet-100K consolidates ten public datasets into a standardized, diverse benchmark at roughly COCO scale.

3.1 Current Status

  • High-resolution SAR datasets are expensive and may contain sensitive information, while small and homogeneous collections can constrain geographic, source, band, polarization, and resolution diversity and bias comparisons.Low-resolution imagery is more available but may not show fine details of smaller objects.Source: §3.1 · p. 4

3.2 SARDet-100K

  • SARDet-100K contains 116,598 images and 245,653 instances across six categories, making it comparable in image count to COCO’s 118K images.The six categories are Aircraft, Ship, Car, Bridge, Tank, and Harbor; the dataset combines ten publicly available datasets and standardizes splits, resolutions, and annotations.Source: §3.2 · pp. 4–5
  • The dataset’s scale and diversity are intended to represent real-world multi-source SAR detection and provide training and evaluation resources.The paper describes it as the first large-scale SAR object detection dataset comparable in size to COCO.Source: §3.2 · p. 5

4 Multi-Stage with Filter Augmentation Pretraining Framework

The MSFA framework combines filter-augmented input with multi-stage pretraining to address gaps between natural-image pretraining and SAR detection fine-tuning. Its two designs target data input, domain transition, and model migration.

4 Multi-Stage with Filter Augmentation Pretraining Framework

  • MSFA combines Filter Augmented Input and multi-stage pretraining to address transfer gaps from data input, domain transition, and model migration.The framework is presented as a response to domain and model gaps in prior pretraining approaches.Source: §4 · p. 5, §4.1.2 · p. 7

4.1 Filter Augmented Input

  • Filter Augmented Input concatenates the original grayscale SAR image with handcrafted filter features to create the detection-model input.The framework casts inputs from heterogeneous pixel space into a homogeneous filter-augmented feature space.Source: §4.1 · pp. 5–6

4.1.1 Multi-stage pretrain

  • Multi-stage pretraining adds optical remote-sensing detection pretraining between ImageNet backbone pretraining and SAR detection fine-tuning.The optical dataset shares object shapes, scales, and categories with downstream SAR datasets and is described as a bridge between natural-image and SAR object distributions.Source: §4.1.1 · p. 6

4.1.2 MSFA

  • MSFA integrates Filter Augmented Input with multi-stage pretraining to bridge domain and modality gaps between natural-image pretraining and SAR detection fine-tuning.The second-stage detection pretraining also trains the whole detection framework, rather than only the backbone.Source: §4.1.2 · p. 7

5 Experiments and Analysis

Experiments report that MSFA improves SAR detection across pretraining strategies, detector frameworks, and backbones. It also achieves state-of-the-art results on SSDD and HRSID.

5.1 Filter Augmented Input

  • Handcrafted filter features improve detection and reduce ImageNet–SARDet-100K distribution gaps; WST performs best, while adding multiple filters yields no significant further gain.Source: §5.1 · p. 7, Table 3 · p. 7, Table 4 · p. 7

5.2 Multi-stage Pretrain

  • Two-stage pretraining improves over baseline strategies; larger DOTA outperforms DIOR, and pretraining the entire framework beats backbone-only pretraining.Source: §5.2 · p. 8, Table 5 · p. 8

5.3 Generalizability of MSFA

  • MSFA improves performance across single-stage, two-stage, and end2end detectors and diverse backbones, with stable gains as backbone size increases.Source: §5.3 · p. 8, Figure 4 · p. 9
  • MSFA is designed to integrate into most existing models without modifications.Source: §5.3 · p. 8

5.4 Comparison with SOTAs

  • 97.9% mAP@50 on SSDD and 83.7% mAP@50 on HRSID set new state-of-the-art results; MSFA is the only open-sourced compared SAR method.Source: Table 6 · p. 9, §5.4 · p. 9

6 Limitaion and Future Work

The paper is limited to supervised pretraining and leaves specialized method designs for future work.

  • The paper studies supervised pretraining only and identifies semi-supervised, weakly supervised, and unsupervised domain transfer as directions for future exploration.Source: §6 · p. 10
  • The method discussion does not explore intricate, specialized designs, which the authors leave for future work.Source: §6 · p. 10

7 Conclusion

The paper introduces SARDet-100k and MSFA for large-scale SAR object detection. It reports improved detection performance, state-of-the-art results on prior benchmarks, and generalizability across models.

  • SARDet-100k contains over 116K images across six categories for SAR object detection research.Source: §7 · p. 10
  • MSFA is a pretraining framework proposed to bridge domain and model gaps between pretraining and finetuning in SAR detection.Source: §7 · p. 10
  • MSFA improves SAR detection performance, sets new state-of-the-art results on previous benchmarks, and generalizes across various models.Source: §7 · p. 10

A.1 SARDet-100K Dataset Visualization

The appendix visualizes representative SARDet-100K samples from six object categories.

A.2 Standardization.

SARDet-100K standardizes collected SAR detection datasets by aligning their splits, image sizes, and annotation formats. The process preserves existing splits where available and prepares oversized images for detection without shrinking targets excessively.

A.2 Standardization.

  • Ten collected datasets undergo split, image-slicing, and annotation-format standardization to create a consistent SARDet-100K pipeline.The dataset overview illustrates these steps in Figure S6(a).Source: §A.2 · p. 16, Figure S6 · p. 17
  • When a source dataset lacks predefined splits, it is divided into train, validation, and test sets at an 8:1:1 ratio.Existing predefined splits are retained.Source: §A.2 · p. 16
  • Images larger than 1000 × 1000 are sliced because resizing could make targets extremely small; AIR SARShip 1, MSAR, and SAR-AIRcraft use 512 × 512 patches with 200-pixel overlap.Figure S6(a) depicts image slicing as part of the standardization process.Source: §A.2 · pp. 16–17, Figure S6 · p. 17
  • All annotations are converted to COCO format, making the merged dataset compatible with popular open-source detection frameworks without extra manual preprocessing.Figure S6(a) includes annotation-format unification in the pipeline.Source: §A.2 · p. 17, Figure S6 · p. 17

A.3 Handcrafted Feature Descriptors

The supplementary experiments describe handcrafted SAR feature descriptors and filter-augmented input, then assess MSFA across models and datasets. They report gains under a similar training budget, broader applicability, and remaining failure cases for difficult imagery and categories.

A.3 Handcrafted Feature Descriptors

  • HOG encodes local structure and shape from gradient orientations; Canny detects significant edges, Haar-like features use predefined templates, and WST captures low- and high-frequency information across scales.The reviewed descriptors include GRE alongside HOG, Canny, Haar-like features, and WST.Source: §A.3 · p. 18

A.4 Multi-Stage Filter-Augmentation

  • Filter-augmented input concatenates the original single-channel grayscale SAR image with a filter-augmentation representation.Figure S8 illustrates this input, and Table S7 compares filter-augmented inputs using Faster R-CNN with ResNet-50.Source: §A.4 · p. 18, Figure S8 · p. 19, Table S7 · p. 19

A.5.1 Main results on SARDet-100K.

  • MSFA significantly enhances performance across single-stage, two-stage, and query-based frameworks, and across ResNets, ConvNets, VANs, and Swin backbones.Tables S8 and S9 report the framework and backbone experiments, respectively.Source: Table S8 · p. 19, Table S9 · p. 20, §A.5.1 · p. 20
  • 1.7% higher mAP on the SARDet-100K test set is achieved by MSFA pretraining versus direct finetuning with the same total epochs and slightly fewer iterations.The comparison uses 12 epochs of MSFA pretraining on DOTA followed by SARDet-100K finetuning, against 36 epochs of direct SARDet-100K finetuning; MSFA pretraining is a one-time effort and its models can be reused across SAR datasets.Source: Table S11 · p. 21, §A.5.1 · p. 21

A.5.2 Comparison of SARDet-100K with other datasets.

  • 8.4% separates the weakest and strongest models on SARDet-100K, compared with 4.1% on SSDD and 4.3% on HRSID.The authors report that model performance on SARDet-100K has not saturated, while larger models continue to benefit there; on SSDD and HRSID, some larger models underperform smaller ones due to overfitting.Source: §A.5.2 · p. 21, Figure S9 · p. 22

A.6 Implementaion Details

  • The second pretraining stage uses DOTA as its main dataset, with DIOR comparison experiments; DOTA images are multiscale-rescaled and cropped into overlapping 1024 × 1024 patches.Finetuning uses 12 epochs for SARDet-100K, SSDD, and HRSID; experiments use MMPretrain and MMDetection on eight RTX-3090 GPUs.Source: §A.6 · p. 22

A.8 Failure Scenarios

  • Failure cases include incorrect classification in detail-poor scenes, missed small dense objects, and localization difficulty; relatively low category performance occurs for small tanks, bridges, and variable-appearance airplanes.The paper focuses on pretraining-to-finetuning domain and model gaps, while noting compatibility with networks designed for these challenging scenarios; Tank has an average area of 461 pixels.Source: §A.8 · p. 23

A.7 Detection Result Visualizations

  • MSFA reduces missed detections, false detections, and localization errors compared with traditional ImageNet backbone pretraining in the visualized examples.Figure S10 presents the comparison.Source: §A.7 · p. 22, Figure S10 · p. 24
Loading 2403.06534v3…