Edinburgh Ceilidh Overhead Video Data

This web page contains video data and ground truth for 16 dances with two different dance patterns. The style of dancing is inspired by Scottish Ceilidh dancing, but the dances shown here are original patterns inspired by two chemical processes. This data was recorded March 7, 2016 at the School of Informatics, University of Edinburgh, as organized by Lewis Hou (Science Ceilidh) and Robert Fisher, and with high resolution filmed by Robin Mair. The dancers were volunteers from the Science Ceilidh and the New Scotland Scottish Country Dance Society and the filming was funded by the Royal Society of Chemistry.

This dataset is interesting because there are very few video analysis datasets where there is highly structured behavior. In this case, the basic moves of all of the dancers are prescribed; however, the timing and positioning can vary by dancer, time and instance of the dance.

The data is acquired using a ceiling camera as a set of 640x480 frames captured at about an average of 8 FPS. Typical frames from the two dances are shown here, along an annotation of the dancers and a `background' frame where no dancers are in the performance area.

Frame from first dance style	Frame from second dance style
Identified dancers in ground truth	Background frame with no dancers

The two dances can be described by a state model, and we have ground-truthed the position of each dancer and their state in each video frame. The ground-truth data can be found below. The format of each ground-truth file is below.

There is some additional useful data:

• The `background' data frame shown above. The dancers largely stay within the space bounded by the 4 white dots.
• The state model and a diagram of the state model for the first dance.
• The state model and part 1 of the diagram of the state model and part 2 of the diagram of the state model for the second dance.
• J. Aizeboje in his MSc project (J. Aizeboje, "Ceilidh dance recognition from an overhead camera", MSc thesis (Informatics), Univ. of Edinburgh, 2016) implemented a CRF in a preliminary study to recognize the state of each dancer in each dance. He achieved an average cross-validated accuracy (over 16 videos) on state label recognition of 96.8% for dance 1 and 94.4% for dance 2 (allowing a ± 5 frame tolerance on labeling). when using the ground truth target positions.
• There is also some unprocessed high resolution video here (1.3 Gb) that can be used with appropriate citations (see below).

There are 16 dances in 2 styles that can be downloaded from the table below, which contains the individual frames, and AVI of the dance, and a ground-truth file labeling the position and current state of each dancer. This tar file contains all of the files from the table (408 Mb). This low resolution video and ground truth can be used with appropriate citations (see below).

Each ground truth datafile contains three arrays:

1. dance_state(N) - the average state (1-M) of the dance for each frame of N in the given dance.
2. dancer_state(10,N) - the current state of each of 10 dancers in the current frame of N in the given dance.
3. dancer_position(10,2,N) - the current (column,row) of each of 10 dancers in the current frame of N in the given dance.

The dance_state label codes are: 1..M for the stages in the dance as seen in the state diagram (e.g. dance1_final_arrows.png). The dancer_state codes describe the immediate state of the dancer at the current frame relative to their contribution to the full dance. For example, for dance style 1, dance 1, dancer 1, frame 80, the dancer_state is 3. Looking at dance1.xls, we see that in this dancer's state 3 the dancer is 'twirling'. The full set of actions is: {standing, walk up, walk down, walk left, walk right, twirling}.

Acknowledgments

Use of the low resolution videos and ground truth: Any use of the video data is under the Attribution-NonCommercial-ShareAlike (aka CC BY-NC-SA) license. Public use of the videos should include this acknowledgment: "We thank the University of Edinburgh for the use of the low resolution video and ground truth data. See J. Aizeboje, Ceilidh dance recognition from an overhead camera, MSc thesis (Informatics), Univ. of Edinburgh, 2016 for more details."

Use of the high resolution videos: Any use of the video data is under the Attribution-NonCommercial-ShareAlike (aka CC BY-NC-SA) license. Public use of the videos should include this acknowledgment: "We thank the Science Ceilidh & Robin Mair, funded by the Royal Society of Chemistry, and the University of Edinburgh for the use of the high resolution video recordings."

Contact

Email: Robert Fisher at rbf -a-t- inf.ed.ac.uk.