Lessons
Each one starts from a recognisable picture, then runs the mechanism on it.
New here? Start with Your First Kumita Adventure
Find and box what is in a picture. Classical window checks and learned detectors are alternative routes, not steps of one.
22 lessons in 8 topic groups (17 free, 5 coming soon · Premium)
Curriculum map
Hover a node for its name · click to open · drag to pan
- needs first
- alternative route
- related
- free
- coming soon · Premium (rebuild)
- coming soon · Premium
- topic group (click to expand)
The curriculum
Lessons
22 lessons on this path
Start with our free lessons. Explore what’s coming next.
22 catalogue lessons shown
New to Kumita
Start here
A short narrated tour: choose a path, read the graph, control a lesson, then rebuild what you learned. Open to everyone.
Feature extraction
Reading a picture with filters
Slide a small grid of weights over a picture to find structure, then shrink what it found.
- Free
Slide the Kernel
Feature extraction
A small grid of weights, one patch at a time, builds a feature map.
Lesson 01 - Free
Keep the strongest response
Feature extraction
Watch a watermelon become a feature map, then shrink the map by keeping the maximum of every two-by-two region, and see what that discards.
Lesson 04
Image representation & preprocessing
Turning a picture into numbers
How colour, shape and scale become the exact values a model takes in.
Neural network fundamentals
Layers of a network
Weighted sums as neurons, and why a network needs a bend between its layers.
- Free
Linear Layer
Neural network fundamentals
Turn real watermelon pixel values into weighted sums: multiply, add, add a bias, see the same arithmetic as a neuron, then run several units as a layer.
Lesson 08 - Free
Why Layers Need a Bend
Neural network fundamentals
Start from a two-switch hallway lamp that no single straight line can classify, then see how a ReLU between two layers makes it solvable.
Lesson 09
Classification outputs
From numbers to class scores
Give a layer one unit per class, read its outputs as class scores, turn them into shares that add up to one, and pick one label.
- Free
Logits
Classification outputs
Give a linear layer one unit per class and read its outputs as raw class scores for the watermelon pixel: negative or above one, not probabilities.
Lesson 10 - Free
Softmax
Classification outputs
Turn the watermelon pixel's raw class scores into positive shares that add up to one: exponentiate, build one shared total, divide.
Lesson 11 - Free
Prediction
Classification outputs
Turn class scores for the watermelon picture into one predicted label with argmax, and judge it against a separately supplied reference label.
Lesson 12
Convolutional neural networks
A whole convolutional network
Follow one picture through every operation, from pixels to class shares.
Object detection
Finding objects
Search a picture, check cheaply first, fit a box, predict every box in one pass, and decide when a detector should say yes.
- Free
Search the picture
Object detection
Move one trained crop classifier across a larger picture.
Lesson 14 - Free
Haar-like features
Object detection
Compare balanced brightness regions before a full examination.
Lesson 15 - Free
Cascades
Object detection
Cheap checks reject most windows before the expensive classifier runs.
Lesson 16 - Free
Make the box fit
Object detection
Describe where the watermelon is with a box, adjust a search window with four numbers, and measure the fit with IoU.
Lesson 17 - Free
YOLO: from features to detections
Object detection
Follow one pixel-art table through a simplified YOLOv3 head: shared features, 140 grid predictions, one decoded box and its scores, a score threshold and class-aware overlap suppression.
Lesson 18 - Free
When should a detector say yes?
Object detection
Move a confidence threshold over seven scored watermelon boxes, check the accepted ones against separate reference boxes, and count true positives, false alarms and misses.
Lesson 19 - Free
Average precision
Object detection
Rank a detector's boxes across three pictures, match each to a reference, follow precision and recall down the list, and measure the area under the precision envelope.
Lesson 20
Coming next
Pictures as numbers
Coming next
Filters and local structure
Coming next
Building classifiers
Coming next