ANN Benchmark results on VGG

VGG Dataset : https://www.robots.ox.ac.uk/~vgg/data/vgg_face2/

ArcFace normalized feature vectors.

Vectors type: float64
Train Vectors: (2701775, 512)
Test Vectors: (10000, 512)
Train Labels: 8631
Test Labels: 5748
Untrained Labels:  0

head(df)

Accuracy

Query Time vs accuracy-Rank@1

In addition to the scatter plot ( transparent points ) we also draw lines connecting points on the Pareto frontier. I.e. the set of optimal of runs that Minimize Query Time and Maximize Accuracy.

Query Time vs accuracy-Rank@10

Index Build Time vs accuracy-Rank@1

The scatter plot shows all the evaluated runs, while the lines show the optimal set of runs.

Showing pareto set of points optimizing R@1 x Build Time. The resulting set is shown in table below.

Index Size vs accuracy-Rank@1

Like in the previous plot we only show the pareto-optimal set of points with high R@1 and low Index Size

Recall

Query Time vs Recall

Index Build Time vs Recall

Index Size vs Recall