Image Rotation Angle Estimation

Try the interactive demo

Predicts the rotation angle of an image using the Circular Gaussian Distribution (CGD) method with a MambaOut Base backbone.

The model outputs a probability distribution over 360 angle bins (1 degree resolution) and extracts the predicted angle via argmax. It handles the full 360 degree range with no boundary discontinuities.

Available Checkpoints

Checkpoint	Dataset	MAE	Median Error
`cgd_mambaout_base_coco2017.ckpt`	COCO 2017	2.84 deg	0.55 deg
`cgd_mambaout_base_coco2014.ckpt`	COCO 2014	3.71 deg	0.68 deg

Usage

The inference code (model_cgd.py, architectures.py, rotation_utils.py) is included in this repo.

from huggingface_hub import snapshot_download

# Download inference code
snapshot_download(
    repo_id="maxwoe/image-rotation-angle-estimation",
    allow_patterns=["*.py", "*.json"],
    local_dir=".",
)

# Load model (defaults to COCO 2017 checkpoint)
from model_cgd import CGDAngleEstimation
model = CGDAngleEstimation.from_pretrained("maxwoe/image-rotation-angle-estimation")

# Or load a specific checkpoint by filename
# model = CGDAngleEstimation.from_pretrained(
#     "maxwoe/image-rotation-angle-estimation",
#     model_name="cgd_mambaout_base_coco2014.ckpt",
# )

# Predict rotation angle
from PIL import Image
image = Image.open("your_image.jpg")
angle = model.predict_angle(image)
print(f"Predicted rotation: {angle:.2f} degrees")

Evaluation Results (COCO 2017, 5 seeds)

Metric	Value
MAE	2.84 deg
Median Error	0.55 deg
RMSE	8.45 deg
P90 Error	3.54 deg
P95 Error	12.00 deg
Accuracy at 2 deg	90.2%
Accuracy at 5 deg	97.5%
Accuracy at 10 deg	98.1%

Model Details

Method: Circular Gaussian Distribution (CGD) — 360 bins, sigma = 6.0 degrees
Backbone: MambaOut Base (mambaout_base.in1k), pretrained on ImageNet-1K
Input size: 224 x 224 pixels
Output: Probability distribution over 360 angle bins, converted to angle via argmax
Loss: KL Divergence with soft Gaussian labels
Optimizer: AdamW with ReduceLROnPlateau scheduler

License

MIT

Downloads last month: 141