Research map / Detection and segmentation / Open-set and few-shot detection
SARDet-100K: Towards Open-Source Benchmark and ToolKit for Large-Scale SAR Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces SARDet-100K, a COCO-scale multi-class SAR detection benchmark merged from 10 public datasets, and MSFA, a multi-stage pretraining framework with handcrafted filter-augmented input (WST) and an optical remote sensing detection bridge stage.
Problem
SAR object detection suffers from small, mono-category public datasets and a domain and model gap when transferring ImageNet-pretrained backbones to SAR detectors, and source code is often unavailable.
Approach
Concatenates handcrafted filter features (notably Wavelet Scattering Transform) with the grayscale SAR image as input, and adds a second detection-pretraining stage on optical remote sensing data (DOTA) that initializes the whole detector rather than only the backbone.
Key results
- MSFA with Faster R-CNN and VAN-B reaches 97.9 mAP50 on SSDD and 83.7 mAP50 on HRSID, reported as new state of the art and the only open-source method among compared SAR detectors.
- On SARDet-100K with Faster R-CNN and ResNet-50, MSFA raises mAP from 49.0 (ImageNet pretrain) to 51.1; MSFA improves all 12 tested detector/backbone combinations.
- Under a matched compute budget (36 epochs), MSFA pretraining yields 54.5 mAP versus 52.8 for direct ImageNet finetuning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply MSFA pretraining instead of standard ImageNet backbone pretraining | Faster R-CNN + VAN-B · SSDD | mAP@50: 92.9 -> 97.9 | Helped |
| Apply MSFA pretraining instead of standard ImageNet backbone pretraining | Faster R-CNN + VAN-B · HRSID | mAP@50: 81.8 -> 83.7 | Helped |
| Add WST features to SAR input (ablation) | Faster R-CNN + ResNet-50 · SARDet-100K test | mAP: 50.2 -> 51.1 | Helped |
| Add Canny, HOG, or Haar filter features to SAR input (ablation) | Faster R-CNN + ResNet-50 · SARDet-100K test | mAP: 50.2 -> 50.7; 50.2 -> 50.7; 50.2 -> 50.6 | Helped |
| Add GRE features to SAR input (ablation) | Faster R-CNN + ResNet-50 · SARDet-100K test | mAP: 50.2 -> 50.6 | Helped |
| Add HOG and Haar features alongside WST (ablation) | Faster R-CNN + ResNet-50 · SARDet-100K test | mAP: 51.1 -> 51.1 | No clear effect |
| Add a second-stage detection pretraining stage on DOTA (ablation) | Faster R-CNN + ResNet-50 · SARDet-100K test | mAP: 49.0 -> 50.2 | Helped |
| Use DOTA rather than DIOR for second-stage pretraining (ablation) | Faster R-CNN + ResNet-50 · SARDet-100K test | mAP: 49.5 -> 50.2 | Helped |
| Pretrain the entire detection framework rather than only the backbone (ablation) | Faster R-CNN + ResNet-50 · SARDet-100K test | mAP: 49.3 -> 50.2 | Helped |
| Apply MSFA instead of ImageNet backbone pretraining | Faster R-CNN with multiple detection frameworks · SARDet-100K test | mAP: 49.0 -> 51.1; 51.1 -> 53.9; 48.8 -> 51.5; 47.4 -> 49.0; 49.8 -> 53.7; 46.5 -> 48.5; 31.8 -> 47.2; 50.0 -> 51.3; 38.1 -> | Helped |
| Apply MSFA instead of ImageNet backbone pretraining | Faster R-CNN with ResNet, ConvNext, VAN, and Swin backbones · SARDet-100K test | mAP: 49.0 -> 51.1; 51.2 -> 52.0; 51.9 -> 52.4; 53.2 -> 54.8; 54.2 -> 55.4; 55.1 -> 56.4; 45.8 -> 47.6; 49.5 -> 51.5; 53.5 -> | Helped |
| Use MSFA pretraining and finetuning rather than direct finetuning | Faster R-CNN · SARDet-100K test | mAP: 52.8 -> 54.5 | Helped |
Surprising findings
- Using multiple handcrafted filter features together (Hog+Haar+WST) gives no further gain over WST alone.
- On small SAR datasets (SSDD, HRSID), larger backbones overfit and underperform smaller ones, whereas on SARDet-100K performance keeps improving with model size.
Limitations
- Scope limited to supervised pretraining; semi-, weakly- or unsupervised domain transfer not explored.
- Failure cases include small, densely packed objects and low-quality imagery; tank and bridge categories perform poorly.
- Method is described as simple and generic without specialized designs for SAR-specific challenges.
Experimental setup
Datasets: ImageNet-1K DOTA DIOR SARDet-100K SSDD HRSID AIR-SARShip MSAR SADD SAR-AIRcraft ShipDataset OGSOD
Benchmarks: SARDet-100K SSDD HRSID DOTA DIOR
Metrics: mAP mAP@50 mAP@75 AP small/medium/large
Baselines: Faster R-CNN Cascade R-CNN RetinaNet GFL FCOS DETR Deformable DETR CenterNet++ CRTransSar SARATR-X
Builds on: ImageNet pretraining Faster R-CNN Wavelet Scattering Transform MMDetection
Closest papers by meaning
- Dual Information Purification for Lightweight SAR Object Detection (AAAI 2025 · similarity 0.87)
- Unleashing Channel Potential: Space-Frequency Selection Convolution for SAR Object Detection (CVPR 2024 · similarity 0.87)
- RSAR: Restricted State Angle Resolver and Rotated SAR Benchmark (CVPR 2025 · similarity 0.86)
- FUSAR-GPT: A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery (CVPR 2026 · similarity 0.86)
- Benchmarking Object Detectors under Real-World Distribution Shifts in Satellite Imagery (CVPR 2025 · similarity 0.85)
- Statistically Principled Deep Learning for SAR Image Segmentation (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.