This new version contains the same images and the same ground truth as the original (http://smartdoc.univ-lr.fr/), but in a format which makes training and testing algorithms easier. We also tried to better identify the tasks researchers can test their methods against using this dataset. Regarding the format, video files were converted to series of images, and ground truth files were converted to a CSV file. Image for the models were also added.
A Python wrapper to facilitate data loading is also available at https://github.com/jchazalon/smartdoc15-ch1-pywrapper.
**If you use Python, we recommend you check the wrapper's documentation directly to use this dataset.**
Tasks
Here are the 3 main tasks which you can test your methods against using this dataset:
1. **Segmentation**: this is the original task.
Inputs are video frames, and expected output is a composed of the coordinated of the four corners of the document image in each frame (top left, bottom left, bottom right and top right).
The evaluation is performed by computing the intersection over union ("IoU" or also "Jaccard index") of the expected document region and the found region. The tricky thing is that the coordinates are projected to the document referential in order to allow comparisons between different frames and different document models.
The original evaluation code is available at https://github.com/jchazalon/smartdoc15-ch1-eval, and the Python wrapper also contains an implementation using the new data format.
2. **Model classification**: this is a new task.
Inputs are video frames, and expected output is the identifier of the document model represented in each frame.
There are 30 models named "datasheet001", "datasheet002", ..., "tax005".
The evaluation is performed as any multi-class classification task.
3. **Model type classification**: this is a new task.
Inputs are video frames, and expected output is the identifier of the document model **type** represented in each frame.
There are 6 models types, each having 5 members, named "datasheet", "letter", "magazine", "paper", "patent" and "tax".
The evaluation is performed as any multi-class classification task.
We also provide model images to allow researchers to experiment methods with more prior knowledge on the expected document images.
Extra files
1. `frames.tar.gz`: gzipped tar archive containing **video frames and their metadata** (ground truth for each task): it can be downloaded from the "Releases" tab at https://github.com/jchazalon/smartdoc15-ch1-dataset/releases
2. `models.tar.gz`: gzipped tar archive containing **model images** (original and variants): it can be downloaded from the "Releases" tab at https://github.com/jchazalon/smartdoc15-ch1-dataset/releases