Loading page
About this tool
Detect and count objects in images using TensorFlow.js COCO-SSD. Supports 80 COCO object categories with confidence scoring and visualization.
AI Object Counter runs the COCO-SSD detector on TensorFlow.js inside your browser to find and count the 80 everyday object classes in the COCO dataset — people, cars, bottles, chairs, dogs, laptops and the rest — drawing a labelled bounding box around each one and a per-class tally beside it. The model is run once at a 20% confidence floor and the results are re-filtered live by a slider, so you can raise or lower the threshold and watch the count change without re-detecting. Nothing is uploaded: the model weights are fetched once and the image is processed on a canvas in the tab.
Open AI Object Counter on AltFTool — it loads instantly in your browser.
Drop an image on the uploader or click to browse (JPG, PNG, WEBP or HEIC, max 20MB); you can also paste from the clipboard or shoot one with Camera. Detection then starts on its own — 'Loading AI model…' and 'Detecting objects…' show over the picture and the status badge switches from Analyzing to Completed.
Drag Confidence Threshold in Detection Settings (10-95%, 40% to start) to drop weak boxes, and flip Show labels off to leave only the numbered badges on the image. Detected Objects lists every box with its number, class, size and score — reorder it with Sort by Confidence, Size or Order, tap a category chip to filter, and use the eye button to highlight one box on the canvas.
Read the results from the Total Objects, Categories, Avg Confidence and Processing tiles plus Object Breakdown and the AI Insights Quality Score, then save from Export Results: Download Annotated Image writes object-detection.png, Export CSV writes object-detection.csv, Export JSON writes object-detection.json, and Export PDF Report opens a printable report.
Detection happens once at a 20% floor and the confidence slider filters client-side, so changing the threshold is instant instead of a fresh model pass.
Every detection gets a cropped 96x72 thumbnail and its own confidence score, so you can see exactly what the model counted rather than trusting a total.
If WebGL warmup stalls past 45 seconds the loader falls back to the CPU backend and retries, so detection still completes on machines where GPU acceleration misbehaves.
The 80 COCO categories — person, bicycle, car, motorcycle, bus, truck, traffic light, dog, cat, bottle, cup, chair, laptop, cell phone and so on. Anything outside that list, including faces, text, logos and specialised industrial parts, will not be counted no matter how clear it is in the photo.
The default is 40%, which is a reasonable balance for clear photos. Lower it toward the 20% floor when objects are small or partly hidden and you would rather over-count and prune manually; raise it above 60% when a false positive costs more than a miss.
The tool uses the lite_mobilenet_v2 base of COCO-SSD, which trades accuracy for speed and struggles with heavily overlapping, very small, or partly occluded objects. Cropping the image into sections and counting each one usually recovers more than lowering the threshold does.
No. The COCO-SSD weights are downloaded to your browser once, and detection runs there via TensorFlow.js — the image itself is only ever drawn to a local canvas. After the first load, detection works with the network disconnected.