Image Rotation Angle Estimation

Estimate how much an image has been rotated from its upright orientation. No labels needed. Includes pretrained models, an interactive demo, and training code for custom datasets.

Quick Start

Requires Python 3.10+, CUDA 11.8+ recommended. Clone the repo, then:

pip install -r requirements.txt

Download a pretrained checkpoint from the HuggingFace Hub and place it in weights/:

weights/
└── cgd_mambaout_base_coco2017.ckpt

Then launch the demo:

python app.py  # opens Gradio UI at http://localhost:7861

For Python inference, see the HuggingFace model card.

Train on Your Own Images

No annotation files or labels needed. Just a folder of correctly-oriented images. The pipeline applies random rotations during training and learns to predict the applied angle.

python train.py --approach cgd --model-name mambaout_base \
  --train-dir path/to/your/images --validation-split 0.1 \
  --batch-size 16 --mixed-precision \
  --test-dirs path/to/test/images --run-test

Evaluation only

python train.py --approach cgd --model-name mambaout_base \
  --train-dir path/to/your/images --validation-split 0.1 \
  --batch-size 16 \
  --test-dirs path/to/test/images \
  --test-only --test-ckpt path/to/checkpoint.ckpt \
  --test-random-seed 0

For reproducing paper results on DRC-D and COCO, see data/DATASETS.md.

Results

Best results on the DRC-D test set (mean of 5 seeds):

Method	Architecture	MAE (°)
CLS	EfficientViT-B3	1.23
CGD	MambaOut Base	1.24

Transfer to COCO, tested on 1,030 val images (CGD + MambaOut Base):

Training Data	MAE (°)
COCO 2014	3.71
COCO 2017	2.84

Methods

Five circular-aware approaches are implemented and benchmarked. CGD performs best overall. See the paper for a full comparison.

Abbreviation	Approach	Description
DA	Direct Angle	Scalar regression with circular loss
CLS	Classification	Angular binning (360 bins)
UV	Unit Vector	Regression on (cos θ, sin θ)
PSC	Phase-Shifting Coder	Frequency-domain encoding
CGD	Circular Gaussian Distribution	Probabilistic soft targets via KL divergence

Citation

@misc{woehrer2026irae,
  title={Image Rotation Angle Estimation: Comparing Circular-Aware Methods},
  author={Woehrer, Maximilian},
  year={2026},
  eprint={2603.25351},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2603.25351}
}