Friday, August 16, 2019

Kshitij Bakliwal, ESSG Visiting Student 2019-20

We present a new algorithm for recognition that uses mass transport distances and amplitude (filtered brightness) errors as inputs of a convolutional neural network that learns to recognize images in Siamese-twin and triplet-loss framework. The net result of this study is an approximately 90% recognition rate on a Gecko dataset in an individual animal identification task..  Our results confirm earlier findings on the value of deformation features (Yang and Ravela, ICCV 09), and additionally suggest that deep learning may need to be substantially primed for effective performance. In particular, local features based approaches failed all together on this dataset and the combination of  l2, transport and CNN appears to give outstanding performance even with very shallow networks.