The convolutional network depth is crucial to accurate large-scale image recognition. In this work, we thoroughly evaluate the networks with increasing depth using the architecture with quite small (3×3) convolution filters. The prior-art configurations can be improved significantly after the depth is pushed to 16–19 weight layers. The comparison with the convolution networks of other convolution filter architectures verifies the effectiveness of the proposed network for large-scale image recognition. In addition, the network verification is conducted with some other data sets to avoid the inherent bias of training data sets. As a result, the most advanced results can be obtained from these data sets.