Python 2.7
Matplotlib
Numpy and Scipy
scikit-learn
Mlpy
It contains 2 folders PoolVideos and Synthetic. The Pool videos are divided into 9 classes and the Synthetic videos are divided into 5 classes as mentioned in the paper.
The Trajectories folder contains two sub-folders Pool and Synthetic, which contain the trajectories extracted from the Pool and Synthetic videos, respectively. The trajectories were extracted using functions available in the OpenCV Library and an open-source version of the Hungarian Algorithm. The trajectories from each video are stored in separate files and are provided in the following format —
Trajectory Index # Starting Frame # X_1 Y_1 X_2 Y_2 . . . X_n Y_n
The UserInput folder contains 2 sub-folders Pool and Synthetic each of which contain the online sketches collected from 25 different users. The data was collected using a Nexus tablet. Each user was shown 10 Pool and 10 Synthetic videos and then asked to sketch the most salient motions from the video which she/he could remember. Each file in Pool/Synthetic folder corresponds to one user containing all the trajectories from all the videos that she/he could remember. The data is stored in the following format —
Video Label # T_1 # T_2 # X_1 Y_1 P_1 X_2 Y_2 P_2 . . . X n Y n P n
where T_1 and T_2 are approximate starting and ending points of the trajectory as remembered by the user normalized to [0, 100]. (X, Y ) and P are the coordinates and pressure values for the sketch respectively.
This folder contains 2 sub-folders Pool and Synthetic, which contain the different sets of features described in the paper. In each file the features are stored in the following format.
For Trajectories, Video Name # Trajectory Label # Starting Frame # Feature String
For User Sketches, User Name # Video Label # Flag # Feature String
Please refer to our paper for a detailed description of the features.
The script RunRetrieval.py needs to be called from the terminal as
python RunRetrieval.py Path/To/Features
Replace ’Path To Features’ with ../Features/Pool/ or ../Features/Synthetic/, accordingly.
The code prints the Mean Reciprocal Rank at the terminal and displays three plots for Precision-Recall, Top-k Accuracy and histogram of reciprocal ranks.
@article{ghosal2014sketch,
title={A Sketch-Based Approach To Video Retrieval Using Qualita-
tive Features},
author={Ghosal, Koustav and Namboodiri, Anoop},
year={2014}
}