Computer Vision Visualizations
Computer vision starts with pixels, but its outputs can mean different things. An edge detector measures local brightness change. An image classifier scores possible labels for one supplied crop. Keeping those contracts separate helps you tell a useful intermediate signal from a prediction about an object.
What These Lessons Cover
| Lesson | Input and output | Experiment |
|---|---|---|
| Edge Detection | A synthetic grayscale grid → Sobel strengths → a thresholded mask | Reverse contrast without changing strength, then test noise and weak edges. |
| Pooling & Feature Maps | Separate feature maps → smaller per-channel summaries | Compare max and average, then move a response across a window boundary. |
| Image Classification | One image crop → a real classifier's 1,000 class scores | Change crop size and position separately; inspect when the top label changes. |
| Object Detection | One photograph → class, score, and box records | Change score and IoU thresholds; separate duplicate boxes from missed objects. |
A Recommended Learning Order
Start with Edge Detection. Calculate a flat patch and a patch crossing a brightness step. Their responses differ because of neighboring values, even when both regions contain bright pixels. Reverse the step and distinguish a signed gradient from its magnitude. The noise scenario establishes the limit: a local change does not necessarily mark an object boundary.
Continue with CNN in the neural-network collection if you want to connect local filters to feature maps and pooling. Its filter editor lets you change what a kernel responds to. The Sobel lesson fixes its kernels so you can isolate polarity and threshold behavior.
Use Pooling & Feature Maps to examine the cost of spatial compression. Keep the same computed feature maps while changing max to average or increasing the window size. A one-pixel shift can disappear from the pooled output; a window-boundary crossing exposes the limit of that tolerance. Each filter's channel stays separate throughout.
Then open Image Classification. It uses recorded SqueezeNet inferences over one photograph. Shrinking a crop and shifting that crop are separate controlled changes. The prediction belongs to the pixels actually supplied to the model, so changing the crop can change the answer even though the original photo stays the same.
Follow with Object Detection. A photo can contain two instances of the same class, so each detection also needs a location. The real two-cat example produces overlapping candidates. Adjust the score and NMS IoU thresholds independently to see why a stricter score filter can miss a cat while a more permissive overlap threshold leaves a duplicate.
Where These Lessons Sit
For direct input manipulation and live inference, use the MNIST Digit Classifier. To understand the class-score normalization behind classification, continue with Softmax & Cross-Entropy.
All Computer Vision Visualizations
Choose a lesson below to inspect its input, play the computation, and test a controlled change.