Generic Sketch-Based Retrieval Learned without Drawing a Single Sketch

09/11/2017

∙

by   Filip Radenovic, et al.

∙

0

∙

share

We cast the sketch-based retrieval as edge-map matching. A shared convolutional network is trained to extract descriptors from edge maps and sketches, which are treated as a special case of edge maps. The network is fine-tuned solely from edge maps of landmark images. The training images are acquired in a fully unsupervised manner from 3D landmark models obtained by an automated structure-from-motion pipeline. The proposed method achieves the state-of-the-art results on a standard benchmark. On two other fine-grained sketch-based retrieval benchmarks, it performs on par with or comes just after the method specifically designed for the dataset.