OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks

OverFeat presents an integrated framework for applying convolutional networks to classification, localization, and detection at once. It shows how a multiscale and sliding-window approach can be efficiently implemented inside a ConvNet, and introduces a novel deep learning approach to localization that learns to predict object boundaries, then accumulates bounding boxes rather than suppressing them to increase detection confidence, with different tasks learned simultaneously using a single shared network.

This integrated framework won the localization task of the ImageNet Large Scale Visual Recognition Challenge 2013 (ILSVRC2013) and obtained very competitive results for detection and classification, later establishing a new state of the art for detection in post-competition work. The authors also released a feature extractor from their best model, called OverFeat, which became a widely used tool for the vision community.