Computer Vision News - November 2016

62 Computer Vision News Tool Results: Conclusion: For many of the 20 objects, t-SNE accurately represents the one-dimensional manifold of viewpoints as a closed loop. For objects possessing similar characteristics from the front and from the back, t-SNE distorts the loop so that the front and back are mapped to nearby points. The four toy car images in the COIL-20 data set are aligned, capturing the similarities between the equivalent elements of all four images. Dataset-3: Dataset with 2 clusters in the 2D plane (most left in the figure below) - For clarity, the two clusters are color coded. The 5 other diagrams show results for five different perplexity values. Results: Conclusion: In his original paper, van der Maaten suggested perplexity values to be in the range (5 - 50). There are indeed two perplexity values in that range (30, 50) which produce a good result in terms of classification (separation of the groups). In terms of topology this range can sometimes be misleading, as we shall see next. Tool