Image Orientation Detection Model

This model identifies the orientation of images, determining whether they are rotated 0°, 90°, 180°, or 270°. It uses a fine-tuned ResNet152 architecture to automatically detect the correct orientation with high accuracy.

Overview

The Image Orientation Detection model solves a common problem in image processing pipelines: automatically determining and correcting image orientation. Whether you’re building an OCR system, processing scanned documents, or organizing a photo collection, having properly oriented images is essential.

This model has been trained on a diverse dataset of images to recognize the natural orientation based on visual cues such as: - Text orientation - Horizon lines - Portrait/landscape recognition - Faces and object positioning

How It Works

The model uses a fine-tuned ResNet152 architecture converted to ONNX format for optimal performance. When an image is submitted, it:

1. Preprocesses the image (resizes and normalizes)
2. Runs inference through the ONNX model
3. Returns the predicted orientation angle (0°, 90°, 180°, or 270°)

For improved accuracy, the model can optionally use a rotation averaging technique that evaluates the image from multiple rotated perspectives and combines the results.

Usage

Basic Usage

import replicate

output = replicate.run(
    "grace-raper/resnet-rot:latest",
    input={"image": open("path/to/your_image.jpg", "rb")}
)

print(f"The image is rotated {output['best_angle']} degrees.")

Advanced Options

output = replicate.run(
    "grace-raper/resnet-rot:latest",
    input={
        "image": open("path/to/your_image.jpg", "rb"),
        "return_probabilities": True,
        "use_rotation_averaging": True
    }
)

print(f"Best angle: {output['best_angle']} degrees")
print(f"Confidence scores: {output['probabilities']}")

API Reference

Input Parameters

Output Format

Standard output:

{
  "best_angle": 90
}

With return_probabilities=true:

{
  "best_angle": 90,
  "probabilities": {
    "0": 0.021,
    "90": 0.965,
    "180": 0.008,
    "270": 0.006
  }
}

Use Cases

• Document Processing: Automatically correct the orientation of scanned documents before OCR
• Photo Organization: Ensure all images in a collection have the correct orientation
• Content Moderation: Preprocess user-uploaded images to ensure consistent orientation
• Data Preparation: Normalize image orientation for machine learning datasets

Model Details

• Architecture: ResNet152 converted to ONNX format
• Input Size: 224x224 pixels (images are automatically resized)
• Model Size: ~60MB
• Runtime: GPU-optimized, with CPU fallback

Limitations

• Performance may vary with extremely unusual images or images without clear orientation cues
• Very small images or images with uniform patterns may be challenging to orient correctly
• The model detects orientation in 90-degree increments only (0°, 90°, 180°, 270°)

Feedback and Contributions

For issues, questions, or feedback about this model, please open an issue in the GitHub repository.

License

This model uses weights from the ResNet-Ixion project and is available for research and commercial use under the terms of the Apache 2.0 license.