Shelf Item Counter Dashboard PoC

I need a compact proof of concept that can take a clear, front-facing photo of a retail shelf, detect every individual item it sees, and then display how many items sit on each shelf row inside a small web dashboard. You will receive a set of high-resolution images shot straight in front of the fixtures—no angled or video frames for this phase—so the model can stay focused on straightforward detection. The objective is simple: if the photo shows four rows, the dashboard should show four numbers, one per row, representing the total count of items the vision model picks up. No brand recognition or empty-space alerts are required yet; raw counting accuracy is the only metric I will review. I am open to the computer-vision stack you prefer (YOLO v5/8, Detectron2, TensorFlow, PyTorch, OpenCV, etc.) as long as the pipeline is reproducible. A lightweight API or script that accepts an image and returns per-row counts is fine; the dashboard can be a minimal Flask, Streamlit, or React front end—just something I can load locally and test with a handful of images. Deliverables • Notebook or script that trains/loads the detection model • Inference routine that groups detections by shelf row and outputs counts • Simple dashboard displaying row-level totals for an uploaded image • Setup notes so I can run everything on my own machine (Linux or Windows) I will consider the PoC successful when I can drop an unseen shelf photo into the tool and reliably see row-level item totals within seconds.