What a detector produces
An object detector takes an image and returns a set of predictions: for each object it finds, a bounding box (where), a class label (what — "person," "vehicle," "bicycle"), and a confidence score (how sure the model is). Modern detectors do this in a single forward pass and can run many times a second on suitable hardware.
That's genuinely useful, and it's also the entire scope of what a detector does. It has no concept of the previous frame, no memory of what it saw a second ago, and no built-in notion of whether a detection matters.
What confidence scores do and don't tell you
A confidence score is the model's own estimate of correctness, calibrated against whatever data it was trained and evaluated on. It is not a guarantee, and it doesn't automatically transfer to a new camera, a new angle, or lighting the model never saw during training. Treating a threshold like "0.8 confidence" as an absolute quality bar, independent of the actual deployment conditions, is one of the most common sources of surprise in production systems.
Why detectors miss things and imagine things
No detector is perfect, and the ways it fails are predictable once you know what to look for:
- Missed detections — occlusion, unusual poses, poor lighting, or objects that simply don't resemble the training data closely enough.
- False positives — shapes or textures that resemble the target class closely enough to fool the model; a common one in surveillance contexts is a mannequin, a poster, or a reflection registering as a person.
- Flicker — the same real object detected in one frame, missed in the next, detected again a frame later — not because anything changed, but because the model's confidence hovered right at the threshold.
These aren't signs of a bad model. They're the expected behavior of a per-frame classifier operating at the edge of its training distribution, which every real deployment eventually does.
Why a better detector usually isn't the whole fix
It's tempting to respond to a noisy system by chasing a higher-accuracy model. That helps at the margin, but most of the noise operators actually experience — duplicate alerts for one person, objects that vanish and reappear, false alarms from a single unlucky frame — comes from asking a detector to do a job it isn't built for: maintaining identity and significance over time. That's tracking's job, not detection's.
The reliable pattern treats detection as a confirmation step, not the whole system: something else (usually motion) proposes that a region deserves attention, and detection verifies and classifies what's actually there — see Motion Detection for the other half of that pairing.
Detection vs classification vs perception
| Task | Answers |
|---|---|
| Classification | "What is this image, overall?" |
| Object detection | "What objects are in this image, and where?" |
| Perception | "What's happening in this scene, over time, and does it matter?" |
See Computer Vision vs Perception for the fuller version of this distinction.
Where Vision Lab fits
In Vision Lab, detection never creates a track on its own — it confirms and elevates tracks that motion has already proposed and sustained. That ordering (motion proposes, semantics confirms) is measured against frozen baselines specifically to check that adding semantic confirmation improves precision without silently changing what the system was already finding. It's the engineering foundation behind Vision Lab Studio, Vision Box, Cabin Cam, and Spy Catcher.
Frequently asked questions
What is object detection?
Object detection is the task of locating objects within an image and classifying what they are, typically output as a bounding box plus a class label and a confidence score.
What does a confidence score actually mean?
It's the model's own estimate of how likely its prediction is correct, not a guarantee. A 0.9 confidence score means the model is usually right at that threshold on its training/evaluation distribution — it says nothing about a specific novel frame in a specific new deployment.
Why do detectors miss objects or produce false positives?
Detectors are trained on finite datasets and generalize imperfectly to conditions that differ from training data — unusual lighting, occlusion, uncommon camera angles, objects that resemble the target class. No detector is perfect across every real-world condition it will encounter.
Is a higher-accuracy detector always the fix for a noisy system?
Not usually. Many production noise problems come from asking a per-frame detector to do a job it can't do alone — maintain identity over time, filter momentary flicker, decide significance. A better detector helps at the margin; the structural fix is usually adding tracking and confirmation logic around it.
What's the difference between detection and classification?
Classification assigns a label to an entire image. Detection locates one or more objects within an image and classifies each one individually, with a position (bounding box) attached.
Where does detection fit in a full perception pipeline?
Detection is typically the confirmation stage that runs on regions a motion or proposal stage has already flagged — verifying and classifying what's there, rather than scanning every pixel of every frame from scratch.
← Back to Knowledge