Fisher Vector Face Descriptor

Karen Simonyan, Omkar Parkhi, Andrea Vedaldi and Andrew Zisserman

Overview

This page contains the download links for the source code for learning and computing the Fisher Vector Face (FVF) descriptor, described in [1]. We also release an extensive set of pre-computed data packages, which can be used to exactly reproduce the results reported in [1].

The FVF descriptors are learnt and evaluated on the Labeled Faces in the Wild (LFW) dataset [2]. We release the learnt descriptor models and pre-computed descriptors for the following LFW evaluation settings:

1. LFW-Unrestricted setting, where the original (unaligned) LFW images are aligned using the method of [3]. Our method achieves 93.03% verification accuracy using joint low-rank metric and similarity learning [1], and 91.32% using a low-rank metric only (which can be seen as discriminative dimensionality reduction).
2. LFW-Unrestricted setting, where unaligned Viola-Jones face detections in the original LFW images are used. Our method achieves 90.68% accuracy using low-rank metric learning. We hope that such pre-learnt low-dimensional descriptors of unaligned face images can find their application beyond the LFW benchmark.
3. LFW-restricted setting, where the pre-aligned LFW-funneled images are used. In this case, our method achieves 87.47% accuracy [1] using diagonal pseudo-metric learning.

For each of the three settings above and the ten dataset splits of LFW, we release:

• the source code for face image pre-processing (alignment and/or cropping) as used in [1];
• Fisher vector computation models (PCA projection for dense SIFT features and a GMM codebook);
• discriminatively learnt face verification models (Setting 1: joint low-rank metric and similarity, and low-rank metric alone; Setting 2: low-rank metric; Setting 3: diagonal pseudo-metric);
• high-dimensional Fisher vector descriptors and low-dimensional projected Fisher vector descriptors (Settings 1 and 2);
• face verification scores.

Downloads

• README.txt: Instructions
• face_desc_src.tar.gz (85.6 KiB): Face descriptor learning, computation, and evaluation source code
• shared.tar (350.1 MiB): Shared data for all evaluation settings (training/testing dataset splits, etc.)
• 1_data.tar (1.3 GiB): Data for setting 1 (LFW-Unrestricted, LFW images aligned using [3])
• 2_data.tar (744.1 MiB): Data for setting 2 (LFW-Unrestricted, unaligned Viola-Jones face detections in LFW)
• 3_data.tar (317.8 MiB): Data for setting 3 (LFW-restricted, cropped LFW-funneled images)

FV descriptors for setting 1

• 1_FV_split1.tar (6.3 GiB): High-dimensional FV face descriptors for split 1 of aligned LFW
• 1_FV_split2.tar (6.3 GiB): High-dimensional FV face descriptors for split 2 of aligned LFW
• 1_FV_split3.tar (6.3 GiB): High-dimensional FV face descriptors for split 3 of aligned LFW
• 1_FV_split4.tar (6.3 GiB): High-dimensional FV face descriptors for split 4 of aligned LFW
• 1_FV_split5.tar (6.3 GiB): High-dimensional FV face descriptors for split 5 of aligned LFW
• 1_FV_split6.tar (6.3 GiB): High-dimensional FV face descriptors for split 6 of aligned LFW
• 1_FV_split7.tar (6.3 GiB): High-dimensional FV face descriptors for split 7 of aligned LFW
• 1_FV_split8.tar (6.3 GiB): High-dimensional FV face descriptors for split 8 of aligned LFW
• 1_FV_split9.tar (6.3 GiB): High-dimensional FV face descriptors for split 9 of aligned LFW
• 1_FV_split10.tar (6.3 GiB): High-dimensional FV face descriptors for split 10 of aligned LFW

Relevant Publications

Acknowledgements

This work was supported by the ERC grant VisRec no. 228180.