Computer Vision Engineer – Visual Product Recognition & Image Retrieval POC

Hi Kashif. I’m developing a visual product recognition POC for a commerce platform. We will initially have approximately 50 sneaker SKUs, with multiple reference images per SKU. A user will upload a real-world smartphone photo of a sneaker, potentially with background clutter, different angles, lighting and partial occlusion. Before discussing the full project, I’d like to understand how you would approach this specific problem. Please briefly explain: 1. How would you detect/isolate the sneaker? 2. How would you generate visual embeddings and retrieve the correct SKU? 3. Which models would you benchmark first (for example CLIP/SigLIP or alternatives)? 4. Would you use FAISS, pgvector or another retrieval approach, and why? 5. How would you combine multiple reference images belonging to the same SKU? 6. How would you distinguish visually similar sneaker models/colorways? 7. How would you handle a sneaker that does not exist in the catalog? 8. How would you measure Top-1 and Top-3 accuracy? 9. Would segmentation necessarily improve retrieval, or would you benchmark both approaches? 10. Have you previously built an image-retrieval, visual-search or product-matching system? If yes, please share the most relevant example. Also(image retrieval / visual product search / metric learning / CLIP-SigLIP / FAISS-pgvector / SKU matching????) At this stage I’m only evaluating the technical approach. Please also give me a rough estimate of the time and cost you would expect for a small POC, not a full application.