Guest blog post by Andrew Brown (PhD researcher), Ernesto Coto (Research Software Engineer) and Andrew Zisserman (Professor) of the Visual Geometry Group, Department of Engineering Science, University of Oxford, and BL Labs Public Award Runner-up for Research, 2019. Posted on their behalf by Mahendra Mahey, Manager of BL Labs.
In this work, we automatically identify and label (tag) people in large video archives without the need for any manual annotation or supervision. The project was carried out with the British Library on a sample of 106 videos from their “Television and radio news” archive; a large collection of news programs from the last 10 years. This archive serves as an important and fascinating resource for researchers and the general public alike. However, the sheer scale of the data, coupled with a lack of relevant metadata, makes indexing, analysing and navigating this content an increasingly difficult task. Relying on human annotation is no longer feasible, and without an effective way to navigate these videos, this bank of knowledge is largely inaccessible.
As users, we are typically interested in human-centric queries such as:
- “When did Jeremy Corbyn first appear in a Newsnight episode?” or
- “Show me all of the times when Hugh Grant and Shirley Williams appeared together.
Currently this is nigh on impossible without trawling through hundreds of hours of content.
We posed the following research question:
Is it possible to enable automatic person-search capabilities such as this in the archive, without the need for any manual supervision or labelling?
The answer is “yes”, and the method is described next.
The basic unit which enables person labelling in videos is the face-track; a group of consecutive face detections within a shot that correspond to the same identity. Face-tracks are extracted from all of the videos in the archive. The task of labelling the people in the videos is then to assign a label to each one of these extracted face-tracks. The video below gives an example of two face-tracks found in a scene.
Two face-tracks found in British Library digital news footage by Visual Geometry Group - University of Oxford.
Techniques at Our Disposal
The base technology used for this work is a state-of-the-art convolutional neural network (CNN), trained for facial recognition . The CNN extracts feature-vectors (a list of numbers) from face images, which indicate the identity of the depicted person. To label a face-track, the distance between the feature-vector for the face-track, and the feature-vector for a face-image with known identity is computed. The face-track is labelled as depicting that identity if the distance is smaller than a certain threshold (i.e. they match). We also use a speaker recognition CNN  that works in the same way, except it labels speech segments from unknown identities using speech segments from known identities within the video.
Labelling the Face-Tracks
Our method for automatically labelling the people in the video archive is divided into three main stages:
(1) Our first labelling method uses what we term a “celebrity feature-vector bank”, which consists of names of people that are likely to appear in the videos, and their corresponding feature-vectors. The names are automatically sourced from IMDB cast lists for the programmes (the titles of the programmes are freely available in the meta-data). Face-images for each of the names are automatically downloaded from image-search engines. Incorrect face-images and people with no images of themselves on search engines are automatically removed at this stage. We compute the feature-vectors for each identity and add them to the bank alongside the names. The face-tracks from the video archives are then simply labelled by finding matches in the feature-vector bank.
(2) Our second labelling method uses the idea that if a name is spoken, or found displayed in a scene, then that person is likely to be found within that scene. The task is then to automatically determine whether there is a correspondence or not. Text is automatically read from the news videos using Optical Character Recognition (OCR), and speech is automatically transcribed using Automatic Speech Recognition (ASR). Names are identified and they are searched for on image search engines. The top ranked images are downloaded and the feature-vectors are computed from the faces. If any are close enough to the feature-vectors from the face-tracks present in the scene, then that face-track is labelled with that name. The video below details this process for a written name.
Using text or spoken word and face recognition to identify a person in a news clip.
(3) For our third labelling method, we use speaker recognition to identify any non-labelled speaking people. We use the labels from the previous two stages to automatically acquire labelled speech segments from the corresponding labelled face-tracks. For each remaining non-labelled speaking person, we extract the speech feature-vector and compute the distance of it to the feature-vectors of the labelled speech segments. If one is close enough, then the non-labelled speech segment and corresponding face-track is assigned that name. This process manages to label speaking face-tracks with visually challenging faces, e.g. deep in shadow or at an extremely non-frontal pose.
Indexing and Searching Identities
The results of our work can be browsed via a web search engine of our own design. A search bar allows for users to specify the person or group of people that they would like to search for. People’s names are efficiently indexed so that the complete list of names can be filtered as the user types in the search bar. The search results are returned instantly with their associated metadata (programme name, data and time) and can be displayed in multiple ways. The video associated with each search result can be played, visualising the location and the name of all identified people in the video. See the video below for more details. This allows for the archive videos to be easily navigated using person-search, thus opening them up for use by the general public.
Archive videos easily navigated using person-search.
For examples of more of our Computer Vision research and open-source software, visit the Visual Geometry Group website.
This work was supported by the EPSRC Programme Grant Seebibyte EP/M013774/1.
 Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, and Andrew Zisserman. VGGFace2: A dataset for recognising faces across pose and age. In Proc. International Conference on Automatic Face & Gesture Recognition, 2018.
 Joon Son Chung, Arsha Nagrani and Andrew Zisserman. VoxCeleb2: Deep Speaker Recognition. INTERSPEECH, 2018
BL Labs Public Awards 2020
Inspired by this work that uses the British Library's digital archived news footage? Have you done something innovative using the British Library's digital collections and data? Why not consider entering your work for a BL Labs Public Award 2020 and win fame, glory and even a bit of money?
This year's public and staff awards 2020 are open for submission, the deadline for entry for both is Monday 30 November 2020.
Whilst we welcome projects on any use of our digital collections and data (especially in research, artistic, educational and community categories), we are particularly interested in entries in our public awards that have focused on anti-racist work, about the pandemic or that are using computational methods such as the use of Jupyter Notebooks.