BOSS dataset
Purpose of annotations
The aim of the annotations is to have a ground-truth file of the BOSS dataset with the following requirements:- It must point out the actions and interactions occurring in the videos by indicating their starting and ending frames
- It must have enough labels to have a wide range of actions and interactions
- It mustn't have too many labels for each action and interaction is represented enough times in order to allow the training of a classifier
- The file must be easy to understand
Explanations
The labels chosen for actions are: walking, sitting, standing, laying down.The labels chosen for interactions are: fight, handshake, kissing on the cheeks, hugging, helping.
As it can be seen on the videos, many others actions and interactions occur such as kneeling, using a cellphone, harassing, reading a newspaper or dancing. However, those events are not happening enough times or are too difficult to precisely point out to be relevant for the purpose of the annotations file.
The has got two worksheets, one for actions and the other for
the interactions.
The actions work sheet is classified by cameras. In each camera
section, there is one array for each video.
In those arrays, the first column indicates the actions
happening in the video while the first row indicates the persons
in the video. The persons are numbered according to their order
of entrance. A person is labelled as entering when half of their
body is visible, it's labelled as leaving when there is less
than half of there body is visible.
When an action is detected, its starting frame number is
indicated in the "start x" column in the corresponding row. When
this action ends, its ending frame number is indicated in the
"end x" column in the corresponding row. As soon as a new action
occurs, its starting and ending frames numbers are annotated in
the same columns as the previous action if the new action is
placed under the previous one. If not, it's placed in a new pair
of columns.
This way, the flow of actions for a person must be read from top
to bottom then left to right, as suggested by the red arrows on
the following picture.
For instance, on the example above, person 1 does action 1 then 2, then 4,then 3, and at last 5. Meanwhile, the person 2 does action 1 then 3 two times before doing action 1 one last time.
For interactions, things are simpler.
As for actions, interactions are indicated on the first row.
For each interaction the number of people involved is indicated,
who they are, and the starting and ending frames as well as
actions.
This way, interactions are indicated in the same fashion as
actions.
