The dataset for this competition contains the vectors of key point coordinates that were estimated using the BlazePose GHUM 3D model. Each vector represents the position of points on a shopper’s body in a particular frame of their shopping session. If the model could not estimate the key point’s location with high confidence, [0.0, 0.0] was used as a missing value for that key point’s location.

Frames are grouped together into groups of 30 consecutive frames (all frames in a group have the same ID in the ID column in the provided datasets). Therefore, you will need to convert the provided datasets into the correct shape before you can train your sequence model. The shape of your dataset after processing should be (number of samples, number of frames, number of features). See the starter notebook for some guidance on how to process the training data (Train.csv) for prediction. The test data (Test.csv) can be processed similarly.

Furthermore, the training target (Train Target.csv) is kept in a separate CSV file as there is only one label per 30 frame sequence. This needs to be kept in mind when building a model.

For this competition, please submit your predicted probabilities of the observations in the test set belonging to class “1”.

Files available for download:

• Train.csv - This is the dataset that you will use to train your model.
• Train Target.csv - This is the target that should be used to train your model
• Test.csv - resembles Train.csv. This is the dataset on which you will apply your model.
• Data Dictionary.csv - This file contains more information on each of the provided features (key point locations).

Consult the figure below for a better idea of where these key points are located on the human body: