Iljung Kwak, Jay Guo, Adam Hantman, David Kriegman, and Kristin Branson
This dataset contains videos and features of mice reaching and eating a food pellet, recorded with two different view points. The MouseReachDataSetVideos_v1.0.zip contains a videos folder with two subfolders: flow and rgb. Each subfolder contains a front and side folder for each view point. The flow folder contains videos of optical flow stored in the same way as I3D. The rgb folder contains 3 color channel videos, where each channel value is equal. Videos are saved as color frames so that it could be used by existing CNN's designed for color videos.
The MouseReachDataSetHDF_v1.0.zip file contains hdf5 files with HoG+HOF features and labels. For each train/test split we have a train and test HDF5 file. Each split hdf5 file has the following keys: date, exp_names, exps, label_names, mice. The exp_names key contains a list of all the experiment names and can be used to index into the exps dataset. For example, indexing hdf5['exps'][hdf5['exp_names'][0]] will access the features and labels for the first video, or more completely hdf5['exps'][hdf5['exp_names'][0]]['labels'] for labels and hdf5['exps'][hdf5['exp_names'][0]]['hoghof'] for hoghof features. The labels are stored as (number of labels) x (number of frames).
This data set is described in Detecting the Starting Frame of Actions in Video.
Videos
Features and Labels