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°	0.55°
`cgd_mambaout_base_coco2014.ckpt`	COCO 2014	3.71°	0.68°

Usage

Download the inference code from this Hub repo (model_cgd.py, architectures.py, rotation_utils.py), then:

python

from model_cgd import CGDAngleEstimation
from PIL import Image

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

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

image = Image.open("your_image.jpg")
angle = model.predict_angle(image)
print(f"Predicted rotation: {angle:.1f}°")

predict_angle accepts a PIL Image, numpy array, or file path.

Evaluation Results (COCO 2017, 5 seeds)

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

Model Details

Method: Circular Gaussian Distribution (CGD), 360 bins, sigma = 6.0°
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