Computer Vision: See → Structure

Classic computer vision turns pixels into edges, depth, poses, boxes, and features. Modern computer vision does the same job with learned encoders (CNNs, Vision Transformers). The output is measurements about the world — geometry, identity, motion — not “pretty pixels.”

The job

flowchart LR
  Pixels["Photos / video"] --> Enc["Vision encoder"]
  Enc --> Struct["Structure:<br/>boxes, depth, pose,<br/>features, embeddings"]

What “structure” means

• Where things are (boxes, masks, depth)
• What they are (classes, identities)
• How they move (tracks, optical flow)
• How they relate (scene graphs, embeddings you can compare)

These are facts the rest of a system can act on: planners, agents, 3D reconstructors, captioners.

Contrast

System	Arrow	Product
Computer vision	pixels → meaning	measurements / embeddings
Diffusion	meaning or noise → pixels	new observations
Gaussian splatting	multi-view photos → editable 3D	a renderable scene

Vision lifts the world into a shared space. Generators and 3D carriers put something back you can see.

Run

File	What it does
edge_measure.py	Synthetic 24×24 scene → Sobel edges → bounding box + centroid (stdlib only)

python computer-vision/edge_measure.py
python computer-vision/edge_measure.py --save computer-vision/out_edges.ppm

WHAT YOU JUST SAW

Pixels became numbers about the world: edge count, gradient strength, box size, centroid. Nothing here invents new pixels — it reads structure out of an image. Contrast that with diffusion, which starts from noise and writes pixels into place.