This archive contains the RWTH-Weather-Phoenix 2014 Multisigner continuous sign language recognition corpus.
It is released under non-commercial cc 4.0 license with attribution (see attachment)

If you use this data in your research, please cite:

O. Koller, J. Forster, and H. Ney. Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers. Computer Vision and Image Understanding, volume 141, pages 108-125, December 2015.

If you use the automatic annotations, please additionally cite:

Koller, Zargaran, Ney. "Re-Sign: Re-Aligned End-to-End Sequence Modeling with Deep Recurrent CNN-HMMs" in CVPR 2017, Honululu, Hawaii, USA.



The multisigner corpus contains 9 signers and has been recorded on the broadcastnews channel.

phoenix-2014-multisigner
├── annotations 
│   │
│   ├── automatic -> this contains the CNN-LSTM-HMM hybrid alignments to train a new system using framelabels
│   └── manual -> this contains the corpus files
│
├── evaluation -> this contains an evaluation script. make sure to have a compiled version of the NIST sclite tools in your path. call: ./evaluatePhoenix2014.sh example-hypothesis-dev.ctm dev
│
├── features
│   │
│   ├── fullFrame-210x260px -> resolution of 210x260 pixels, but they are distorted due to transmission channel particularities, to undistort stretch images to 210x300
│   │   ├── dev
│   │   ├── test
│   │   └── train
│   │
│   └── trackedRightHand-92x132px -> we further provide dumped rectangles containting the right (dominant) hand of the signers. To undistort those scale them to squared size
│       ├── dev
│       ├── test
│       └── train
│
└── models -> we provide caffe models to achieve 27.1 / 26.8 % WER on the dev/test partition of this corpus, we also provide our languagemodel