Description:

We provide two large-scale collections for experiments with the libmsvm package:

1. the trecvid 2015 collections for semantic indexing task and
2. the voc 2012 collections.

Each data-set is divided into the development and test data-sets.

The development data-set is also splitted into two subsets (training and validating) in-order to be used with cross-validations. The descriptor and the labels are provided for all samples in different ways (i.e. separated files, one file, binary and .svm formats).

Trecvid_sin15

• annotations of 346 visual concepts of Trecvid SIN 2013 are provided.
• Full collection 2013
  1. descriptor binary
  2. separated annotations ' Two formats i) binary-labels (-1, 0, +1) and ii) multi-labels {+-1 .. +-346}'
  3. .SVM Format (i.e. annotation + descriptor vector in sparse mode) ' Two formats i) binary-labels and ii) multi-labels'

• Sub collections :Each subset contains a select.txt file, which indicates the keylist samples to be selected from the full collection.
  1. Developement set 2013d: divided into two subsets 2013xand 2013y
  2. Test set 2013z.

voc12

• annotations of 20 visual concepts of Pascal-VOC 2012 are provided.
• Full collection 2012
  1. descriptor binary
  2. separated annotations ' Two formats i) binary-labels (-1, 0, +1) and ii) multi-labels {+-1 .. +-20}'
  3. .SVM Format (i.e. annotation + descriptor vector in sparse mode) ' Two formats i) binary-labels and ii) multi-labels'

• Sub collections :Each subset contains a select.txt file, which indicates the keylist samples to be selected from the full collection.
  1. Developement set Dev-set: divided into two subsets train-set and validation-set
  2. Test set test.