Research map / Detection and segmentation / CNN features
ReDiffDet: Rotation-equivariant Diffusion Model for Oriented 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.
Formulates oriented object detection as a rotation-equivariant denoising diffusion process by representing oriented boxes as 2D Gaussians, with a conditional encoder and aligned decoder (RiRoI Align).
Problem
Applying diffusion-based detectors to oriented object detection is hard due to arbitrary orientations, randomly placed boxes that poorly cover dense objects with extreme aspect ratios, and misalignment between oriented boxes and axis-aligned image features.
Approach
Represent each oriented box as a 2D Gaussian so the diffusion reverse process can be proven rotation-equivariant; add a learned conditional encoder that proposes conditional boxes as diffusion starting points, and an aligned decoder using RiRoI Align to align oriented boxes with features.
Key results
- ReDiffDet reaches 75.45% AP50 on DOTA-v1.0 with ReResNet50 and 80.70% under multi-scale training, versus 60.50% for DiffusionDet-O.
- On DIOR-R, ReDiffDet achieves 68.05% AP50 versus 49.91% for DiffusionDet-O (ablation baseline).
- On DOTA-v1.5 and v2.0, ReDiffDet reaches 68.00% and 55.76% AP50, surpassing ReDet and OrientedFormer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add rotation-equivariant diffusion model (over DiffusionDet baseline) (ablation) | DiffusionDet-O on DIOR-R · DIOR-R | AP50: +2.24 (49.91 to 52.15) | Helped |
| add conditional encoder (conditional boxes) (ablation) | ReDiffDet on DIOR-R · DIOR-R | AP50: +7.87 (52.15 to 60.02) | Helped |
| add aligned decoder (RiRoI Align alignment) (ablation) | ReDiffDet on DIOR-R · DIOR-R | AP50: +2.15 (60.02 to 62.17 with decoder alone on top of encoder setup) | Helped |
| add conditional encoder and aligned decoder together to rotation-equivariant diffusion (ablation) | ReDiffDet on DIOR-R · DIOR-R | AP50: +15.68 over the rotation-equivariant-only baseline (52.15 to 68.05 with all components) | Helped |
| add conditional boxes (conditional encoder) to DiffusionDet | DiffusionDet on COCO · COCO | mAP: +1.8 (43.2 to 45.0) | Helped |
| increase decoder layers from 3 to 4 (vs 5-7) (ablation) | ReDiffDet on DIOR-R · DIOR-R | AP50: peak at 4 layers (68.05); 3 layers 67.74 AP50, 5 layers 67.74 AP50 | Mixed |
| increase number of sampled points k from 8 to 128 (ablation) | ReDiffDet on DIOR-R · DIOR-R | AP50: 68.05 at k=8 down to 67.37 at k=128 | Hurt |
| use scaling factor m=6 vs 1.0 or 8.0 (ablation) | ReDiffDet on DIOR-R · DIOR-R | AP50: 68.05 at m=6 vs 67.26 at m=4 and 66.72 at m=1 | Mixed |
| ReDiffDet with ReResNet50 vs DiffusionDet-O baseline on DOTA-v1.0 | ReDiffDet · DOTA-v1.0 | AP50: +14.95 (60.50 to 75.45) | Helped |
| ReDiffDet vs DiffusionDet-O on DIOR-R small/large-aspect objects | ReDiffDet · DIOR-R | AP50: +18.14 (49.91 to 68.05) | Helped |
| use ReResNet50 backbone vs ResNet50 or Swin-T backbone | ReDiffDet on DOTA-v1.0 · DOTA-v1.0 | AP50: ReR50 75.45 vs R50 76.18 vs Swin-T 76.62 | Mixed |
| condition the box centers using the conditional encoder at inference, when only conditional boxes are used (ablation) | ReDiffDet on DIOR-R · DIOR-R | AP50: conditional boxes reach 67.7 at 1 step, plateau; noisy-only reaches 69.3 after 8 steps | Mixed |
Surprising findings
- Conditional boxes give higher accuracy in the first iteration but accuracy does not improve much with further iterations, since the box distribution differs from training data.
Limitations
- Rotation-equivariance is proven only for the reverse process with DDIM sampling; other sampling strategies are not analyzed.
- Non-rotation-equivariant backbones (ResNet-50, Swin-T) are used without equivariant features, with gains attributed to parameter count.
Experimental setup
Datasets: DIOR-R DOTA-v1.0 DOTA-v1.5 DOTA-v2.0 COCO
Benchmarks: DOTA-v1.0 DOTA-v1.5 DOTA-v2.0 DIOR-R COCO
Metrics: AP50 AP75 AP50:95 mAP FPS parameters
Baselines: DiffusionDet-O Oriented R-CNN RoI Transformer AOPG ARS-DETR OrientedFormer FRED ReDet DCFL
Builds on: DiffusionDet ReDet DDIM MMRotate
Closest papers by meaning
- FRED: Towards a Full Rotation-Equivariance in Aerial Image Object Detection (AAAI 2024 · similarity 0.87)
- R2Det: Exploring Relaxed Rotation Equivariance in 2D Object Detection (ICLR 2025 · similarity 0.87)
- 3DiffTection: 3D Object Detection with Geometry-Aware Diffusion Features (CVPR 2024 · similarity 0.86)
- Rethinking Boundary Discontinuity Problem for Oriented Object Detection (CVPR 2024 · similarity 0.86)
- CamoDiffusion: Camouflaged Object Detection via Conditional Diffusion Models (AAAI 2024 · similarity 0.86)
- DiffuBox: Refining 3D Object Detection with Point Diffusion (NeurIPS 2024 · similarity 0.86)
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.