Visual Inventory Pro — count shelf items from pixels, with a row for every box

1. Source image

Sample seed Items per shelf Pixel noise Damaged items drawn

Drop a shelf photo on the view below, or pick one:

No photo loaded yet. Sample data is in use.

The three sample buttons draw a synthetic shelf with this page and process it as an image, so a run can be checked against a known item count. That generated count is shown in the sidebar and labelled as sample ground truth rather than as a detection result.

2. Detection parameters

Adaptive window (px)31

Threshold offset C12

Min box area (px)60

Closing radius (px)0

Damage edge-var min80

Polarity

reject components touching the frame edge (shelf boards and cut-off items) show rejected shapes

Every control above feeds the same pipeline. Nothing is cached: each change re-thresholds the pixels, re-labels the connected components and re-derives the count, so the number on screen really moves with the parameters.

3. Detection view

source + bounding boxes (drop an image here)

adaptive threshold mask

Green = counted box. Red = counted box flagged as a damage candidate. Grey dashed = component rejected by the current rules (border artefact, too small, or degenerate size).

4. Audit trail

Sort by

#	label	x	y	w	h	area px	mean gray	mean edge	edge var	flag	note

How the detection works

Every stage runs on the canvas pixels in this tab: grayscale, adaptive local-mean threshold, optional morphological closing, 8-connected two-pass labelling with union-find, per-box area, mean luminance and mean Sobel edge, and damage flagging by edge-density variance over a 3x3 cell grid inside each box.

Damage flags are a heuristic shortlist of candidates, not a classifier verdict. Cracks, tears, crushed corners and dents raise local edge variance — and so do printed logos, barcodes, string wrap and hard shadows. Read the flag as the list for a human to confirm, which is exactly what the audit table underneath is for. Every count and every damage flag is a shortlist to check by hand rather than a verdict: a pixel heuristic over-reports.

Counting is a function of the parameters, and the audit table shows every component with the reason it was counted or rejected. Recognition, labelling, pricing, depth and occlusion reasoning belong to a model rather than to this pixel pipeline: a box hidden behind another cannot be counted, and two touching items merge into one component unless closing and the area rules separate them.