Date: Tuesday, May 17 (Main Conference Day 1)
Start Time: 1:30 pm
End Time: 2:00 pm
In this talk, we present a novel approach to improving the accuracy of convolutional neural networks (CNNs) used for classification. Our approach utilizes the confusion matrix of the original CNN on a specific dataset to identify sets of low accuracy classes that resemble each other with respect to the error distribution. Using this information, several shallow networks are generated which operate in parallel with each other and evaluate input frames before the frames reach the original, large CNN. The shallow networks are able to classify the low-accuracy classes more accurately than the original network, while eliminating the need to run the larger network on certain images. Hence, by combining the shallow networks with the original network, accuracy is improved, with virtually no increase in inference time.