Illumination-Robust Segmentation

Easy

Segment an outdoor scene when the light will not hold still. Colour normalisations that ignore brightness, casts, and shadows; equalisation and Otsu so thresholds pick themselves; erosion to clean the mask; IoU, mean IoU, and soft Dice; then the two identities that make the model lightweight: separable-conv cost and BatchNorm folding.

Learning Objectives

→Normalise RGB to chromaticity, gray-world, and a log-chromaticity invariant, and state what each one is invariant to
→Equalise a histogram and choose an Otsu threshold from the between-class variance
→Erode a binary mask with a zero-padded 3x3 window
→Compute IoU and Dice for one mask, mean IoU from a confusion matrix, and the soft Dice loss
→Count the multiply-accumulates of a depthwise-separable layer and fold BatchNorm into a convolution for inference

Coding Problems (11)

Chromaticity: Colour Without Brightness

~8 min· Intro

Gray-World White Balance

~10 min· Intro

Log-Chromaticity Illumination-Invariant Image

~14 min· Easy

Histogram Equalisation

~14 min· Easy

Otsu's Threshold

~16 min· Easy

Binary Erosion with a 3x3 Window

~12 min· Easy

IoU and Dice for a Binary Mask

~10 min· Intro

Mean IoU from a Confusion Matrix

~12 min· Easy

Soft Dice Loss

~12 min· Easy

Depthwise-Separable Convolution Cost

~12 min· Easy

Fold BatchNorm into the Convolution

~16 min· Medium

Illumination-Robust Segmentation

Learning Objectives

Coding Problems (11)

Concept Check