SVTR variable width

[crashilko]

Hello!

I have a question about the ability of the SVTR recognizer to process images of an arbitrary widths.
what should I do if my images, for example, have shape of 32x1000?

[GreatV]

Hello!

SVTR (Spatial Visual Transformer for Recognition) in PaddleOCR is designed to process text images of varying widths, but there are some considerations when handling extremely wide images like 32x1000.

Steps to Handle Arbitrary Widths:

1. Adjust the Model's Input Shape:

   • PaddleOCR typically resizes images to a fixed height (e.g., 32) while keeping the aspect ratio.
   • If your image is 32x1000, ensure that the model can handle such long sequences without excessive compression.

2. Modify the Configuration:

   • Check the rec_image_shape parameter in the model configuration.
   • If it is set to a fixed width (e.g., [3, 32, 320]), you may need to adjust it to allow longer widths, such as [3, 32, None] (if supported).

3. Use a Dynamic Length Model (if available):

   • Some OCR models support dynamic sequence lengths, meaning they can process images without forcibly resizing to a fixed width.
   • If SVTR does not support dynamic width, you may need to resize the width while preserving important details.

4. Consider Splitting the Image:

   • If the model struggles with very wide inputs, an alternative is to split the image into smaller segments and process them sequentially.

Additional Considerations:

• Memory Usage: Longer sequences require more GPU memory. Ensure your hardware can handle the increased computational load.
• Performance Tuning: Experiment with different resizing strategies to find the best trade-off between accuracy and efficiency.

Would you like guidance on modifying the PaddleOCR configuration for this specific use case?

---

Response generated by feifei-bot | chatgpt-4o-latest

[crashilko]

In file predict_rec.py it is stated, that longer images are resized to fixed width. Can I change this parameter and resize with fixed ratio, without retraining the model and without any affects in performance?