Xerox Research Centre Europe (XRCE) computer-vision team

What XRCE is

XRCE (Xerox Research Centre Europe) is Xerox's European research laboratory. Its computer-vision team competed in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC). The challenge's overview paper records that Fisher vector-based methods performed strongly in every ILSVRC from 2010 to 2014, and XRCE's entries belong to that record ^[1]. The team's work centred on Fisher vectors and related image representations: fixed-length encodings of local image content that can be fed to linear classifiers ^[1].

2010: an honorable mention

In 2010 an XRCE team received an honorable mention. Its entry used an improved Fisher vector representation with PCA dimensionality reduction, followed by data compression and a linear SVM ^[1]. The paper credits Perronnin and Dance 2007 for the improved Fisher vector representation that the 2010 entry used ^[1]. The sequence—high-dimensional encoding, dimensionality reduction or compression, then linear classification—recurs in the team's two documented entries.

2011: winning the classification task

The following year XRCE won the classification task. The paper notes that the winning classification entry in 2011 was the 2010 runner-up team XRCE ^[1]. This entry applied high-dimensional image signatures with compression using product quantization and one-vs-all linear SVMs ^[1]. The paper cites Perronnin et al. 2010 for the signatures and Sanchez and Perronnin 2011 for the product quantization used to compress them ^[1]. Compared with 2010, the 2011 system kept the broad structure but used the newer signature and compression methods ^[1].

How the components worked

A Fisher vector describes an image by how its local descriptors deviate from a generative model of descriptor space. Rather than a histogram of visual words, it keeps statistics that capture more detail, at the cost of a much longer vector ^[1]. PCA projects such a vector onto the directions of greatest variance, reducing its length ^[1]. Product quantization goes further: it cuts the vector into subvectors, quantizes each against a small codebook, and stores the resulting short codes, so that many signatures can be kept in memory and compared approximately ^[1]. A one-vs-all linear SVM trains a separate linear classifier for each class, and a test image is assigned the class whose classifier scores highest ^[1]. The combination made rich, high-dimensional encodings practical to use with simple and fast classifiers.

Reach and influence

The paper observes that Fisher vector-based methods evolved over the five years of the challenge and continued performing strongly in every ILSVRC from 2010 to 2014 ^[1]. Their influence extended beyond XRCE. The ISI team that placed second in 2012 image classification used Fisher vectors credited to Sanchez and Perronnin 2011 ^[1]. The 2012 runner-up VGG entry used Fisher vectors together with dense SIFT features, color statistics and a linear SVM ^[1]. These results place XRCE among the groups whose representation work was taken up by other top teams in the challenge ^[1].