perception lab

See when a vision model stops seeing clearly
under blur, low light, motion, and occlusion.

Upload an image or use a webcam. SIGNALWATCH runs COCO-SSD in the browser and records only what the model actually reports: detections, confidence, timing, and failures.

model outputs

detections and confidence are reported only when COCO-SSD emits them

telemetry

stability, variance, persistence, and replay are computed from detection history

input controls

degradation presets alter pixels before inference; they are not model confidence

real source

Claims need source records or ingestion events.

model output only

Confidence and detections come from inference.

unavailable is valid

Missing data stays visible instead of guessed.

no prefilled claims

Protocols do not ship with conclusions.

how to read this lab

Input

The image or webcam frame the model sees.

Stress

Blur, low light, crop, motion, occlusion, or noise added before inference.

Output

Only the boxes and confidence values the browser model actually reports.

Failure

A missing object, unstable class, low confidence, or empty frame.

live perception test

browser inference

1 input

Upload an image or enable webcam.

2 stress

Apply blur, low light, crop, motion, or compression-like noise.

3 infer

Run COCO-SSD locally and inspect only reported detections.

4 export

Open the evidence packet and export JSON after real frames exist.

upload image or enable webcam to run real browser-side detection

idle

1 upload image / real COCO-SSD inference

sample route / none / detections remain model-reported only

sequencenone selected

asset statead hoc input

min framesunbounded

webcam routeidle

input condition control88

mean detection confidenceunavailable

tracking persistenceunavailable

detections 0 / frames 0

observation windowpending

inference cadence0.8s target

continuity transitions0

last frameunavailable

robustness sequence registry

5 protocols / 0 asset-backed / 5 capture-required

Registry entries are capture protocols until real frames are imported. Selecting one applies reproducible degradation metadata; detections and confidence remain model-output only.

calibration fixtures / synthetic geometry only / excluded from dataset registry

2 apply operating stress

blur0

low light100

occlusion12

compression/noise8

crop instability0

motion blur0

temporal confidence trace0 frames

waiting for model outputs

observation window / confidence bars from model outputs / muted break markers indicate dropped detections

windowpending

cadencepending

jitterpending

transitions0

continuity markers

no continuity transition recorded in the current window

class persistence windows

no emitted class has persisted across this window

DETECTION.RELIABILITYunavailable

rolling mean confidence from model outputs

CONFIDENCE.VARIANCEunavailable

inverse variance across recent inference frames

TRACKING.PERSISTENCEunavailable

class continuity between adjacent frames

TEMPORAL.CONSISTENCYunavailable

confidence movement between adjacent frames

PERCEPTION.STABILITYunavailable

stability of detected object count

FRAME.INTEGRITYunavailable

share of recent frames with at least one model detection

OCCLUSION.PRESSUREunavailable

derived from detection disappearance and class continuity loss

4 operational evidence packet

sequencenone

asset statusad hoc

modeupload

presetdefault

frames0

with detections0

empty frames0

mean confidenceunavailable

drop events0

continuity breaks0

persistence windows0

windowpending

cadence jitterpending

No dataset sequence is selected. The export will record this as an ad hoc browser inference run.

No model detections have been emitted yet. The packet remains empty instead of synthesizing evidence.

waiting

No inference frames have been recorded in this observation window.

continuity transitions

no transition recorded

export boundary / model outputs, timestamps, degradation settings, derived notes, and evidence packet only

temporal replay

replays detection history only; no video frames or confidence values are synthesized

waiting for inference history

frame / unavailable

confidence / unavailable

classes / none reported

operational explainability

No inference history yet. Enable webcam or upload an image to begin collecting real model outputs.

provenance / latest explanation derived from detection history and current model outputs

real model readout

no detection reported by model

robustness trace

Baseline detections are captured from the same uploaded image before degradation. The degraded frame is re-run after slider changes; instability is measured only from COCO-SSD outputs.

baseline / unavailable

Model inference remains unavailable until the browser model finishes loading. Confidence values shown here come only from real model outputs.

plain boundary

real outputs only

purpose

This lab shows how detections change when the input gets harder to read. It is not a ranking of models.

data boundary

Detections, confidence, persistence, and replay traces come from browser-side COCO-SSD outputs. Missing outputs stay visible.

deployment surface

Inference runs locally in the browser. The backend does not generate or simulate detections.

case studies

future pose tests

not simulated

pose model disconnected / no keypoints emitted

joint confidenceunavailable

temporal jitterunavailable

keypoint persistenceunavailable

architecture ready / waiting for browser pose backend

pose safety context

Pose stability monitoring is prepared for future MediaPipe or TensorFlow pose estimation. Until a real model is connected, this panel does not display skeletons, confidence, or joint telemetry.

safety-critical context

autonomous vehicles

plain meaning / seeing the road clearly

Perception must remain stable across lighting, motion, partial visibility, and sensor noise before downstream planning can be trusted.

robotics

plain meaning / knowing where objects are

Manipulation and navigation depend on object continuity. Detection collapse can change how a robot estimates reachable space.

industrial automation

plain meaning / working around moving equipment

Factory vision systems need robustness under vibration, compression, glare, and occlusion from moving equipment.

surveillance systems

plain meaning / showing uncertainty to people

Uncertainty should be surfaced clearly; degraded inputs can reduce reliability even when a model appears confident in clean conditions.

medical imaging

plain meaning / avoiding overconfidence

Safety-critical imaging requires explicit evaluation under acquisition artifacts, missing context, and distribution shift.

biomechanical analysis

plain meaning / keeping movement stable over time

Pose and movement systems need temporal stability; jitter or missing keypoints can distort downstream movement signals.

pose tracking systems

plain meaning / not guessing hidden joints

Occlusion, clothing, camera angle, and low light can reduce joint persistence. Confidence should be measured, not assumed.

perception evidence packet shape

schema preview / values require real run

schema

signalwatch.perception.evidence.v1

generatedAt

set when a real run is exported

model

browser-side COCO-SSD when loaded

inferenceBoundary

local browser inference only

observationWindow

derived from actual frame timestamps

frames

empty until frames are observed

detections

empty until model emits detections

continuityTransitions

computed from emitted class history

non-populated export shape

{
  "schema": "signalwatch.perception.evidence.v1",
  "generatedAt": "<real export timestamp>",
  "model": "<loaded browser model>",
  "inferenceBoundary": "real outputs only",
  "frames": [],
  "detections": [],
  "continuityTransitions": []
}

what counts as evidence in this lab

plain language / trust boundary

Real

Observed or sourced

A source link, timestamp, runtime event, or model output exists. The interface can point back to where it came from.

Derived

Calculated from real inputs

A summary, grouping, or trace built from source activity, telemetry, or model-output history. It should still show its inputs.

Conceptual

Explanation, not measurement

A teaching example that explains a risk or system behavior without claiming it happened in a deployed system.

Simulated

Controlled demonstration

A parameter-driven demo. Useful for learning, but not evidence about the outside world unless real inputs are attached.

perception unavailable states

absence is observable / absence is not guessed

model unavailable

perception lab

If browser-side inference cannot load, detections remain unavailable instead of simulated.

no detections emitted

model output

Empty frames are recorded as empty frames; no bounding boxes or confidence values are fabricated.

collector offline

source ingestion

Source health reports offline or delayed states directly through collector telemetry.

insufficient window

evaluation

Temporal claims wait until a real observation window has enough frames or source events.

packet not exported

evidence packet

The interface can show the export schema without pretending an evidence packet exists.

source missing

provenance

A claim without a usable source trace should stay unresolved, not become narrative filler.

perception operating conditions

illustrative photos / not model evidence

These source-attributed photos show visual conditions the lab is built to reason about. They do not contain detection boxes, confidence values, or precomputed outcomes; those are produced only when the browser model runs.

visual role

operational context

source status

externally attributed

evidence boundary

not inference output