ImageNet is a large-scale image dataset organized into a hierarchy of semantic categories. It became a proving ground for whether artificial neural networks could handle visual recognition at scale. section 2.3
Fei-Fei Li began developing ImageNet in 2006, drawing inspiration from Irving Biederman’s theory that recognizable objects combine a limited set of basic shapes. The dataset and accompanying paper appeared in 2009; its first major competition followed in 2010. section 2.3
Early ImageNet competitions favored handcrafted features and SVMs, with only incremental gains. In 2012, AlexNet cut top-5 error from 26% to 15%; the result showed that more data alone was not enough. section 2.3 section 2.5.5
Figure 2.6 The top-5 error rates in the ILSVRC from 2010 to 2017. SIFT and SVM won in 2010 and 2011, but performance plateaued. CNNs won from 2012 to 2017, including AlexNet in 2012, and eventually saturated the benchmark.
The error-rate timeline shows the shift from SIFT-and-SVM winners to CNNs, which won from 2012 through 2017. section 2.5.5
By 2013, nearly all competition entries used CNNs. ImageNet’s benchmark success helped establish deep learning in computer vision and inspired further architectures; later work also applied CNNs to object detection. section 2.5.5
