Large-Scale Knowlege Transfer

[Full set of software and data is available upon request]

Software

Semantic relatedness from Language resource

For now see our previous release at www.d2.mpi-inf.mpg.de/nlp4vision. We will shortly add an updated version.

Large Scale Learning

In our recent work [1] we learn a SVM with MeanSGD. The Matlab/C++ trains several classes in parallel using multi-core hardware, loading the data only once to memory. This software is based on code by Leon Bottou.

meanSGD 1.1

Change log

1.1: Added input check and "-largeArrayDims" for mex compilation.

Problems/Errors

Undefined function or variable 'meanSvmsgd2HingelossOpenmp'

Make sure the binary is on the path and you provide data of the type "single". Alternatively change the C++ code to double.

Image Features

The features are zipped Matlab (mat) files, one per image.

• Fisher
  • training (145 GB)
  • validation (6 GB)
  • test (18 GB)
• LLC
  • training (38 GB)
  • validation (1.5 GB)
  • test (4.5 GB)

Precomputed SVM classifier scores

• 1-vs-all case for test, training and validation set.
  as Matlab matrixes.
  • all (zipped, 19 GB)
    • fisher (training, validation, test)
    • llc (training, validation, test)
    • rgSift (training, validation)

Groundtruth labels

• groundtruth (0.3 MB)
• zero shot
  • class split
    • known (800)
    • unkown (200)

Known Issues

• black-white images as we use rgSift.

Poster

Help & Contact

If you need any help or have any suggestions feel free to contact me.