Research map / Detection and segmentation / CNN features

FRED: Towards a Full Rotation-Equivariance in Aerial Image Object Detection

AAAI 2024 · Detection and segmentation › CNN features · approach: Box regression and loss design · 26 citations (OpenAlex)

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 FRED, a fully rotation-equivariant oriented object detector for aerial images that represents boxes as rotation-equivariant point sets and uses them as deformable-convolution offsets, achieving competitive DOTA results with ~16% of the parameters.

Problem

Oriented object detection in aerial images relies on rotation-invariant features, alignment heuristics, and heavy rotation augmentation, and does not achieve true end-to-end rotation-equivariance; bounding-box orientation prediction is not equivariant.

Approach

Decouple classification (rotation-invariant, via orientation alignment with an equivariant reference vector) from localization (rotation-equivariant point set as vector field), and use the equivariant point set as offsets in rotation-equivariant deformable convolution (RE-DCN) and rotation-invariant DCN (RI-DCN), plus an edge constraint loss to stabilize the reference point.

Key results

• FRED achieves 75.56 mAP on DOTA-v1.0, comparable to state-of-the-art (best anchor-free Oriented RepPoints 75.97).
• FRED achieves 68.30 mAP on DOTA-v1.5, outperforming Oriented RepPoints (66.90) and ReDet (66.86) by about 1.4-1.5 mAP.
• FRED model size is 22.28 MB versus 140.78 MB for Oriented RepPoints and 124.97 MB for ReDet, about 16% of the parameters.

Measured effects reported in the paper

Surprising findings

• Before full convergence, FRED appears to learn relative object poses within a class without direction-specific supervision, a tendency that fades as training converges.
• Ablation shows group pooling alone (73.91) improves over the no-equivariance baseline (73.32) only modestly, and the full model adds about 2.2 mAP over the baseline.

Limitations

• Strict equivariance is only exact for 90-degree rotations on discrete images; strided convolution breaks equivariance unless odd-sized inputs are used.
• Results are single-scale only; the DOTA-v1.0 result is comparable rather than superior to the best anchor-free baseline.
• The directionality-emergence observation is qualitative and weakens in converged models.

Experimental setup

Datasets: DOTA-v1.0 DOTA-v1.5 ImageNet

Benchmarks: DOTA-v1.0 DOTA-v1.5

Metrics: mAP

Baselines: RoI-Transformer FAOD SCRDet Gliding Vertex R3Det S2A-Net DAL KFIoU ReDet Oriented RepPoints CFA RSDet++

Builds on: ReDet Oriented RepPoints RoI Transformer E(2)-CNN MMRotate

Closest papers by meaning

• ReDet: A Rotation-Equivariant Detector for Aerial Object Detection (CVPR 2021 · similarity 0.94)
• ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects (CVPR 2025 · similarity 0.89)
• R2Det: Exploring Relaxed Rotation Equivariance in 2D Object Detection (ICLR 2025 · similarity 0.88)
• ReDiffDet: Rotation-equivariant Diffusion Model for Oriented Object Detection (CVPR 2025 · similarity 0.87)
• Relaxed Rotational Equivariance via G-Biases in Vision (AAAI 2025 · similarity 0.85)
• FBRT-YOLO: Faster and Better for Real-Time Aerial Image Detection (AAAI 2025 · similarity 0.85)