Research map / Detection and segmentation / Open-set and few-shot detection
AeroGen: Enhancing Remote Sensing Object Detection with Diffusion-Driven Data Generation
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.
Proposes AeroGen, a layout-controllable latent diffusion model for horizontal and rotated boxes in remote sensing, plus a filtered end-to-end generative pipeline that synthesizes detection training data.
Problem
Remote sensing object detection suffers from scarce labeled data, and existing generative methods do not support layout control with both horizontal and rotated bounding boxes for generating detection training data.
Approach
A layout-conditioned LDM that encodes 8-coordinate boxes (covering both HBB and OBB) with Fourier encoding and CLIP class embeddings, injects layout via a layout mask attention and dual cross-attention, and adds a DDPM layout-label generator plus CLIP/classifier image filters to augment detection datasets end-to-end.
Key results
- AeroGen reaches FID 38.57 and CAS 76.84 on DIOR versus 41.31 and 63.50 for GLIGEN in layout-to-image generation.
- Adding 50k synthetic images improves DIOR-R mAP from 37.39 to 41.69 (mAP50 60.21 to 64.12) over the real-data baseline.
- Synthetic data raises mAP by 3.7% on DIOR, 4.3% on DIOR-R, and 2.43% on HRSC, with gains of 17.8% on GF and 14.7% on DAM classes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add AeroGen synthetic images to YOLOv8 training | YOLOv8 · DIOR | mAP: 54.22 -> 57.92 | Helped |
| Add AeroGen synthetic images to YOLOv8 training | YOLOv8 · DIOR-R | mAP: 37.39 -> 41.69 | Helped |
| Add AeroGen synthetic images to YOLOv8 training | YOLOv8 · HRSC | mAP: 63.49 -> 65.92 | Helped |
| Use AeroGen instead of the strongest listed baseline, GLIGEN | Layout-to-image generation · DIOR | FID: 41.31 -> 38.57 | Helped |
| Use AeroGen instead of GLIGEN† | Layout-to-image generation · DIOR-R | FID: 48.43 -> 35.07 | Helped |
| Use AeroGen instead of GLIGEN† | Layout-to-image generation · HRSC | FID: 66.69 -> 45.86 | Helped |
| Add Layout Mask Attention (LMA) (ablation) | AeroGen · DIOR image generation | FID: 82.11 -> 40.76 | Helped |
| Add Dual Cross Attention (DCA) to local and global LMA (ablation) | AeroGen · DIOR image generation | FID: 40.76 -> 38.57 | Helped |
| Add diverse layout synthesis and layout-condition filtering (ablation) | AeroGen augmentation pipeline · DIOR-R | mAP: 41.31 -> 41.69 | Helped |
| Add layout-condition filtering (ablation) | AeroGen augmentation pipeline · DIOR-R | mAP: 40.92 -> 41.69 | Helped |
| Add synthetic-data augmentation (ablation) | AeroGen augmentation pipeline · DIOR-R | mAP: 39.62 -> 41.69 | Helped |
| Add semantic-consistency image filtering (ablation) | AeroGen augmentation pipeline · DIOR-R | mAP: 40.27 -> 41.69 | Helped |
Surprising findings
- Plain copy-paste and flip augmentation give smaller gains than AeroGen synthetic data on the same detector.
- Under the HRSC 100-image few-shot setting, adding synthetic data substantially raises mAP (35.20 to 46.23).
Limitations
- Image generation is limited to remote sensing layouts trained per dataset; generalization across datasets is not shown in the main text.
- Filtering thresholds and filter strictness hyperparameters are not characterized in depth.
Experimental setup
Datasets: DIOR DIOR-R HRSC DOTA1.0
Benchmarks: DIOR DIOR-R HRSC DOTA1.0
Metrics: FID CAS YOLO Score mAP mAP50
Baselines: LostGAN ReCo LayoutDiffusion GLIGEN Copy-Paste Oriented R-CNN Redet Faster R-CNN
Builds on: Stable Diffusion GLIGEN LDM Oriented R-CNN YOLOv8
Closest papers by meaning
- SatSynth: Augmenting Image-Mask Pairs through Diffusion Models for Aerial Semantic Segmentation (CVPR 2024 · similarity 0.81)
- DiffusionSat: A Generative Foundation Model for Satellite Imagery (ICLR 2024 · similarity 0.81)
- ROS-SAM: High-Quality Interactive Segmentation for Remote Sensing Moving Object (CVPR 2025 · similarity 0.80)
- Sat2Scene: 3D Urban Scene Generation from Satellite Images with Diffusion (CVPR 2024 · similarity 0.80)
- MMM-RS: A Multi-modal, Multi-GSD, Multi-scene Remote Sensing Dataset and Benchmark for Text-to-Image Generation (NeurIPS 2024 · similarity 0.79)
- DiffuBox: Refining 3D Object Detection with Point Diffusion (NeurIPS 2024 · similarity 0.79)
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.