YOLO: real-time detection, from zero
A 4-lesson mini course: why detection was slow, the single-glance idea, the cleanup math, and the honest trade-offs.
Why detection was slow
Classifiers wearing a detective costume
Before YOLO, object detection reused image classifiers — networks that answer 'what is this picture of?'. To find objects, systems showed the classifier thousands of crops of the same image: sliding windows at every position and scale (DPM), or ~2,000 'this might be something' region proposals (R-CNN). One photo meant thousands of separate neural network evaluations.
Pipelines, not systems
R-CNN's pipeline had separate stages: propose regions, extract features, classify with SVMs, then refine boxes — each trained independently. Errors compounded between stages, tuning was painful, and even the streamlined Faster R-CNN topped out around 7 frames per second. Accurate, but you could not point it at live video.
The convolution foundation
One thing did work brilliantly: convolutional feature extraction — sliding small filters across an image to find edges, textures, and parts (exactly what the CNN Lab below shows). YOLO's bet was that a convolutional backbone could feed a detector that decides everything in one shot, with no proposal stage at all.
Key takeaway
Pre-YOLO detection = thousands of classifier calls glued into fragile multi-stage pipelines. Accurate but far too slow for live video.
Hands-on — experiment before the quiz
CNN Convolution Lab
How computers see: a tiny magnifying glass scanning a picture
Input
Feature map (10×10)
Compares each pixel's left side to its right side. Big difference = a vertical edge lives here. Press play to watch the filter scan the image — this exact multiply-and-add is what every CNN layer does, millions of times.
Analogy: Sliding a small stencil over a photo and asking at every spot: how much does this patch look like my pattern?
Checkpoint
Why was R-CNN-style detection so slow?