As described in the paper, we have annotated a very small datasets with the setting of stride/window as 1 / 6. The annotation targets were created with training_data_serialization/newer_college_labelling_data_genearator.py, which you can take a closer look to see how it works.
Please download the datasets with this link: https://drive.google.com/file/d/1-mFaLKPZSeVgdZWgCx4rym-0sWlURw4u/view
Unzip it and you will find the structure as:
├── 0
│ ├── agreed_labels.txt
│ └── testing
└── 1
├── agreed_labels.txt
└── testing
├── 39_seqid_13823_4490553f59a047d3a344c0e202b0da51_gt5.png
├── ....png
The testing folder contains the exact visualization image with the following information encoded (taking
39_seqid_13823_4490553f59a047d3a344c0e202b0da51_gt5.png as one example):
- id in the sub-testing folder, e.g.
39 - unique sequence id, e.g.
13823, which can help you identify the frame in the raw data - a hash
- psudo ground truth/baseline synced results provided by Newer College, e.g.
5
The agreed_labels.txt reads as:
id,label
0,disagreed
1,disagreed
2,disagreed
3,5
4,disagreed
5,5
6,3
7,3
8,disagreed
....
where the first column is the id in the sub-testing folder, which can be matched to image in the testing folder; the second column shows the annotated results, with one of the following:
- a single or several number(s), which all labelers agree
- disagreed
- unsure
In our test, only the agreed ones are used.
Technically, we don't need the images in the folder at all. However, I still provide it in this way since this shows how the data are annotated and this is also a format that directly works with newer_college_gt_runner.py.
In the meanwhile, I also provide a processed annotation.csv, which contains all information, without requiring the images. This is in case that the download link is broken and someone still would like to use these datasets. But, you will need to make some changes to the newer_college_gt_runner.py to get it work. The annotation.csv is created by the converter.py.