Playground experiment

Computer Vision Mini Lab

Small visual toys around detection, saliency, reconstruction error, masks, and image intuition.

Synthetic image probe

One scene, five model views

A neutral generated image is reused across five views, so the model overlay changes while the underlying scene stays visually stable.

Raw image

Probe readout

Scene overview

Detector focus 0 regions

Mask coverage 0%

Reconstruction error Low

Selected pixel Background

Threshold 0.58 Image noise 0.16

model overlay error / reveal probe

What to notice

Detection compresses an image into boxes and confidence scores. Saliency spreads attention over pixels. Reconstruction error shows where a simple prior fails to explain what is visible.

From pixels to pipeline

The scene is synthetic, but the habits are real: compare views, move thresholds, inspect false certainty, and keep the model output tied to the image that produced it.