Week 03 · lesson

Investigation: Transformations Change the Evidence

System pipeline

From Captured Values to Bounded Vision Evidence

Read from top to bottom. Each arrow means the next stage uses the result or information produced by the stage above it.

Scene Condition

Declare lighting and crop before interpreting stored values.

BASELINENORMAL LIGHT

Pixel Tensor

Represent the supplied teaching image as 6 by 6 coordinates with three channels.

SHAPE6 × 6 × 3

Framework Operation

Apply declared numeric transformations without confusing the library with the model.

MODESRGB · GRAY · EDGES

Feature Measurement

Record selected-pixel values and marked edge locations.

NORMAL @ 9012 / 36

Controlled Test

Change one scene or threshold condition while holding the other settings fixed.

DIM @ 908 / 36

Claim Boundary

Report feature evidence without claiming object identity or understanding.

ACTIONREVISE CLAIM
Read this concept flow as plain text
  1. Scene Condition. Declare lighting and crop before interpreting stored values. BASELINE: NORMAL LIGHT.
  2. Pixel Tensor. Represent the supplied teaching image as 6 by 6 coordinates with three channels. SHAPE: 6 × 6 × 3.
  3. Framework Operation. Apply declared numeric transformations without confusing the library with the model. MODES: RGB · GRAY · EDGES.
  4. Feature Measurement. Record selected-pixel values and marked edge locations. NORMAL @ 90: 12 / 36.
  5. Controlled Test. Change one scene or threshold condition while holding the other settings fixed. DIM @ 90: 8 / 36.
  6. Claim Boundary. Report feature evidence without claiming object identity or understanding. ACTION: REVISE CLAIM.

A computer-vision pipeline often transforms an image before producing a result. A transformation can make a feature easier to measure, but it can also remove information or expose a fragile shortcut.

From framework operation to vision evidence

A framework may place a tensor on a device, run declared operations, and track a computational graph. Those conveniences do not decide whether the input represents the intended scene. A perfectly executed operation on a cropped or poorly lit image still processes incomplete or changed evidence.

Resize changes grid dimensions. Grayscale combines color channels. Thresholding separates values using a cutoff. Edge detection marks locations where neighboring values differ enough.

if neighbor_difference >= threshold:
    edge = 1
else:
    edge = 0

An edge is evidence of a sharp numeric change. It is not an object name.

Controlled comparison

Suppose the supplied normal scene produces 12 marked edge pixels at threshold 90 and the dim scene produces 8. The physical object did not lose four edges. The capture condition changed stored values, which changed the transform.

If you also lower the threshold, two conditions changed. You can no longer isolate which change explains the result. Hold the transform fixed while changing the scene—or hold the scene fixed while changing one transform setting.

Completed comparison

  1. Question: Does dim lighting change the supplied edge evidence?
  2. Hold fixed: 6 × 6 grid, grayscale representation, edge rule, threshold 90.
  3. Change one condition: normal scene to dim scene.
  4. Measure: normal 12/36; dim 8/36.
  5. Difference: 8 - 12 = -4 marked pixels.
  6. Bounded claim: Under the supplied grids and fixed threshold, dim light produced four fewer marked locations. This does not establish how another camera, image, or model behaves.

Investigation

  1. Predict which supplied condition—normal, dim, or cropped—will produce the highest edge count at threshold 90.
  2. Predict whether raising the normal-scene threshold from 90 to 120 will mark more or fewer locations.
  3. Name the independent variable, measurement, and fixed conditions for each comparison.
  4. Preserve both predictions. A failed prediction becomes evidence when the measurement explains the defect.
  5. Write one sentence that uses “edge evidence” without saying “object detected.”

Success means another student can repeat the comparison from your declared settings.

Vocabulary lab

Flip the idea, not just the card

Explain the term before you reveal the back. Then compare your explanation with the definition, example, and warning.

1 / 5
Read all terms without animation
tensor
A multidimensional numeric array used by a framework to represent data and intermediate values. Example: A color image can be represented as height by width by three channels. Do not confuse it with: A tensor is a data structure, not an understanding of the scene.
transformation
A declared operation that changes an image representation. Example: Converting RGB values to grayscale. Do not confuse it with: It changes available evidence, not necessarily the physical scene.
threshold
A cutoff used to place numeric values into different output groups. Example: Mark an edge when a difference reaches 90. Do not confuse it with: A threshold is not proof that one cutoff works in new conditions.
edge
A location where nearby image values change sharply under a declared rule. Example: The boundary between an orange region and gray background. Do not confuse it with: An edge is not an object label.
controlled comparison
A test that changes one declared condition while holding the others fixed. Example: Compare normal and dim scenes at threshold 90. Do not confuse it with: Changing lighting and threshold together is not controlled.