Weak Labels for PersonID in TV series

This is a Matlab implementation of the paper:

---

Improved Weak Labels using Contextual Cues for Person Identification in Videos
Makarand Tapaswi, Martin Bäuml, and Rainer Stiefelhagen
IEEE International Conference on Automatic Face and Gesture Recognition (FG), 2015
Paper download | ShotThreading & SceneDetection code

---

Tested on

Ubuntu 14.04 with Matlab version R2014a - R2015a.

First initialization

The first_init.m script will be called on running startup.m the first time. This will ask you to install some external toolboxes. Please follow the instructions.

---

Example usage

A video structure can be created by calling either
VS = BBT(1, 1); or VS = BUFFY(5, 1)

The main function can be directly invoked with one or multiple videos at once. ft = speaking_face2_wrapper(BBT(1, 1:6));

We include data for 6 episodes each of

• The Big Bang Theory (BBT) season 1, episodes 1..6
• Buffy the Vampire Slayer (BUFFY) season 5, episodes 1..6 as described in the paper.

For any questions about the generation of data, please contact me.

---

External toolboxes

• jsonlab: Matlab JSON interface
• DataHash: Create a hash of parameters for caching
• maximalCliques: Computes maximal cliques
• minmaxk: Min-k Max-k (in MEX!)

Main functions

• speaking_face2_wrapper.m Main function which runs over all episodes
• speaking_face2_prepare_data.m Prepares data, handles whether to use context such as threading and uniqueness
• speaking_face2_optimize_fmincon.m Runs fmincon with the correct options over all cliques
• speaking_face2_fmincon_objfun.m Contains the core objective function used to compute labeling scores for each face track clique

Changelog

• 12-03-2015: A complete working example of the FG 2015 paper