Computer Vision News - February 2018

CIFAR The CNN used for CIFAR consisted of 2 convolutional layers: the first had 32 5×5 filters, the second had 64 5 × 5 filters, the first followed by a Relu activation layer and both followed by max pooling layers, downsampled by a factor of 2. Facial Expression Recognition challenge For the Facial Expression the CNN used an image mirroring layer, similarity transformation layer, two convolutional filtering + pooling stages, followed by a fully connected layer with 3072 hidden units. Source code Results: The author showed that a CNN network with L2-SVM loss function outperformed softmax on 2 popular benchmark classification datasets. Results are as follows: MNIST CIFAR Facial Expression 6 Computer Vision News Research Research ConvNet+Softmax ConvNet+SVM Test error 0.99% 0.87% ConvNet+Softmax ConvNet+SVM Test error 14.0% 11.9% ConvNet+Softmax ConvNet+SVM Accuracy 14.0% 11.9%