---
dataset_info:
- config_name: multisigner
  features:
  - name: id
    dtype: string
  - name: transcription
    dtype: string
  - name: frames
    sequence: image
  splits:
  - name: train
    num_bytes: 35090755574
    num_examples: 5672
  - name: validation
    num_bytes: 3294869318
    num_examples: 540
  - name: test
    num_bytes: 3935898314
    num_examples: 629
  download_size: 43042303939
  dataset_size: 42321523206
- config_name: pre-training
  features:
  - name: id
    dtype: string
  - name: transcription
    dtype: string
  splits:
  - name: train
    num_bytes: 744118
    num_examples: 5672
  - name: validation
    num_bytes: 63848
    num_examples: 540
  - name: test
    num_bytes: 75329
    num_examples: 629
  download_size: 43042303939
  dataset_size: 883295
- config_name: signerindependent
  features:
  - name: id
    dtype: string
  - name: transcription
    dtype: string
  - name: frames
    sequence: image
  splits:
  - name: train
    num_bytes: 26933922764
    num_examples: 4376
  - name: validation
    num_bytes: 720569029
    num_examples: 111
  - name: test
    num_bytes: 1175797903
    num_examples: 180
  download_size: 29320607031
  dataset_size: 28830289696
---
# RWTH-Weather-Phoenix 2014

This archive contains two sets of the RWTH-Weather-Phoenix 2014 corpus

 1. the multisigner set 
 2. the signer independent set.

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.
```

and

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

See README files in subfolders for more information.

### CHANGELOG
 - v1 Aug 20  2016, initial version of the archive. multisigner setup
 - v2 Apr 21  2017, signer independent SI5 subset, added caffe models and automatic frame-alignment
 - v3 Nov 3  2017, added language models and complete set of hyper parameters to reproduce the published results