Object Detection for Dummies Part 2: CNN, DPM and Overfeat

Lilian Weng · 8 years ago

This article explains three classical approaches to object detection: convolutional neural networks (AlexNet, VGG, ResNet) that form the foundation for deep learning vision tasks, the Deformable Parts Model which recognizes objects using part-based graphical models, and OverFeat which integrates classification and bounding box prediction into a single CNN. AlexNet (2012) established the CNN baseline with 5 convolution layers, VGG (2014) extended this to 19 layers using only 3x3 convolutions, and ResNet (2015) scaled to 152 layers using residual blocks. OverFeat pioneers the unified approach by training a CNN for classification then adding a regression network to predict bounding box coordinates at multiple scales, evaluated using mean average precision (mAP) with intersection-over-union threshold typically at 0.5.