Best approach for football shirt image similarity and categorization with manual review threshold?

Beginners
1

Hi everyone,

I’m working on a project to categorize and match football (soccer) shirt images from resale listings with a database of known shirt models. Here’s what I’m trying to achieve:

  1. Compare an input image (from a resale listing) with a database of standard football shirt images.

  2. Identify the most similar shirt model in the database.

  3. If the similarity score is below a certain threshold, flag it for manual review.

Some details:

  • Database size: Currently about 300 unique shirt models, each with 5-10 images (Stock images). About 30-40 items matching the shirt model from listings on resale websites having each 4-5 photos (Manually categorized)

  • Input images: Varied quality, different backgrounds, sometimes worn by people.

  • Need to handle different angles, lighting conditions, etc.

I’m looking for recommendations on the following:

  1. The most efficient and accurate approach for this type of image similarity task.

  2. Suitable models or architectures (I’ve heard about Siamese networks and metric learning, but I’m open to other suggestions).

  3. Any pre-trained models that might be particularly useful for this kind of task.

  4. Techniques to handle the manual review threshold effectively.

I’m comfortable with Python and have basic knowledge of TensorFlow and PyTorch. Any advice, resources, or pointers would be greatly appreciated!

Thanks in advance for your help!