We visualize the retrieval results for SHREC13 dataset in this demo. The window on the left shows the Sketch-Map, which is the 2D PCA projection of the learned CNN feature. Each hexagon represents a number of data points within the region, and one sketch is shown in the center. Different color denotes different data point density in the region. Please use the scroll wheel to zoom in and discover more sketches.
Mouse usage:
As mentioned in the paper, we include the precision-recall curves for the PSB/SBSR here and here.
Source code:
The full source code (ShapeRetrieval-release.tar.gz, 16KB, README.txt) is available for downloading.
Our training code is written in python, based on the Theano framework, and the evaluation codes in Matlab are adapted from the SHREC13 benchmark.
We also provide the processed data (Shrec13-preprocessed-data.tar.gz, 64.4MB) we used for the SHREC13 dataset. To use the data, please uncompress the tar file in the same directory with the code. To keep the data file small, we removed the augmentation data, which can be easily generated with the matlab script included in the code package.