Vision

Perception Beginner Runs on this server Real-time

Draws a box around every object it recognises and reports how confident it is. It knows exactly 80 categories and nothing else — watching it fail on the 81st is the interesting part.

Model: YOLOv8n
Parameters: 3.2M · smallest of the five YOLOv8 sizes
Input: Image
Output: Boxes + labels
Typical time: 0.1–1s
Last run: —

01 Your image

Drop, paste or choose JPG · PNG · WEBP · max 10MB

Samples

02 Confidence threshold

Minimum score 25%

Applied after detection, so moving it is instant — no second request. Detections below the line are discarded, not hidden: the model still produced them.

What confidence means This is the model's own estimate that a box contains an object of that class. It is calibrated against training data, not against reality. A 0.95 on an object type the model has seen a million times means something quite different from a 0.95 on an unusual one.

03 Display

“Boxes” draws the overlay in your browser from returned coordinates, so it responds to the threshold. “Model render” is the annotated frame the server produced at detection time, at the original threshold.

Output Awaiting an image

Nothing detected yet

Add an image on the left, then run the detector. Try the night-concert sample first — it is the more interesting failure.

Detections —

Each detection will appear here with its confidence score. Hover a row to isolate that object in the image.

Why did it detect these? YOLOv8 divides the image into a grid and, in one forward pass, predicts for every cell whether an object centre lies there, the box dimensions around it, and a probability across 80 fixed classes. It has no concept of anything outside those 80 categories. When it sees one, it does not report uncertainty — it reports the nearest class it does know, or nothing at all.

How it works

The image is divided into a grid. For every cell the network predicts, in a single forward pass, whether an object centre falls there, how big it is, and which of 80 classes it belongs to. That one-pass design is what the name means: You Only Look Once.

Input Your image, resized to 640×640

Model One forward pass through a 3.2M-parameter CNN

Filter Non-max suppression removes duplicate boxes

Output Boxes, class labels, confidence scores

Full identifier: Ultralytics YOLOv8-nano
Where it runs: On this server, on CPU
Parameters: 3.2M

Trained on the COCO dataset, which is heavy on everyday Western scenes. Unusual objects, extreme angles, motion blur and heavy occlusion all degrade it sharply.

What to notice

Confidence is not accuracy. A 0.9 means the pattern strongly matched what the model saw in training, not that it is 90% correct.
Raise the threshold and small or partly hidden objects vanish first — those are the detections the model was least sure about.
Anything outside the 80 COCO classes gets forced into the nearest one, or missed entirely. There is no 'I don't know' output.

A note on language

This page says “the model predicts”, not “the model knows”. That is deliberate. None of these systems understand the images or sentences you give them; they map inputs to outputs using patterns fixed at training time. The difference matters most exactly when the output is impressive.