Oct 30, 2019 · newsletter

Fuzzy People

While preparing for a recent edition of the Federated Learning talk I often give at conferences, I encountered this tweet, which includes a demonstration of real-time multi-person segmentation on a smartphone.

Some useful terminology here:

  • Object detection typically refers to identifying and localising objects (such as people!) in an image, and surrounding them with a bounding box.
  • Object segmentation labels each pixel of the image with a class (for instance: “person,” “car,” “cat,” …).

Segmenting multiple people in an image is substantially harder than segmenting a single person, because identifing the parts of the image that belong to each person essentially requires modeling each person’s pose. Without understanding human poses, how would our function-approximating neural network know that an arm around a shoulder belongs to person one, rather than person two?

This technology working on a mobile device in real-time is impressive.

There is good reason to be concerned about the malicious or misguided use of facial recognition technology to invade personal privacy. However, there are many conceivable applications for computer vision in public spaces which do not require incidental privacy violation - monitoring road traffic, for example.

A potential use of the raft of face-swapping apps recently launched is protecting the identity of people in the background of video or photos, while maintaining the realism of the image.

This naturally raises the question: what is private?

If you know me well, you can probably recognise me with an obfuscated face, based on clothes and surroundings. The overt pixellation in the mobile demo appealed to me. Because vision so viscerally connects to our senses, it strikes me as a fascinating arena in which to test theoretical measures of privacy.

I managed to find a few older papers covering this idea - though given it’s import, it feels under-explored.

When in doubt, play.

Thanks to the ready availability of pre-trained models, TensorFlow.js, Observable notebooks, and some free time at a conference, I could reasonably straightforwardly construct a person pixellator, which you can try for yourself here: Fuzzy People.

It can take an input image, such as this:

Brooklyn Bridge with pedestrians (image courtesy of

And output (an attempt at) a more private version:

Brooklyn Bridge with fuzzy pedestrians

Since the picture is really intended to be of the Brooklyn Bridge and Manhattan skyline, we’ve done the pedestrians a favour and hidden their identities.

Clearly, it doesn’t work perfectly.

Some experimentation reveals it to work best when the people are in the foreground, clearly separated, there are only a few of them, and the picture was taken in reasonable lighting and weather conditions.

Nonetheless, I’m impressed by how well it sometimes does work, given that it is running some quite sophisticated models entirely inside a web browser.

The pixellation technique used is simple: for each pixel that the model identifies as a person (the model here is a combination of a bounding box with COCO-SSD, and single person segmentation with BodyPix), replace it with a randomly selected pixel from within an adjustable region. At low noise, where the region the random pixel is selected from is small, people in the images are still human identifiable. As noise is increased, people progressively become less person-like and more static.

Finding a less intrusive - but still highly private - means of masking people from images is left as an exercise to the research community.

The app is extremely early stage work, but I think credit is due to the open source community for providing pre-trained models to enable the creation of something that would have seemed magic a decade ago in a day or so of hacking. Certainly, we are past the point where some elements of computer vision can be considered as commodity: simply import the “detect person” function.

A side-note on the pace of change in this space:

Alas, much of my work is wasted. About a week after I made the notebook work, a multi-person version of BodyPix was released! While I confess to feeling slightly stung by having sunk a few hours into pixel manipulation to combine a bounding box model with a single-person segmentation model that has ultimately proved unnecessary, I’m excited to try out the new model. It certainly brings home the feeling of rapid progress in computer vision technology.

Read more

Nov 21, 2019 · newsletter
Oct 30, 2019 · newsletter

Latest posts

Nov 15, 2020 · post

Representation Learning 101 for Software Engineers

by Victor Dibia · Figure 1: Overview of representation learning methods. TLDR; Good representations of data (e.g., text, images) are critical for solving many tasks (e.g., search or recommendations). Deep representation learning yields state of the art results when used to create these representations. In this article, we review methods for representation learning and walk through an example using pretrained models. Introduction Deep Neural Networks (DNNs) have become a particularly useful tool in building intelligent systems that simplify cognitive tasks for users. more
Jun 22, 2020 · post

How to Explain HuggingFace BERT for Question Answering NLP Models with TF 2.0

by Victor · Given a question and a passage, the task of Question Answering (QA) focuses on identifying the exact span within the passage that answers the question. Figure 1: In this sample, a BERTbase model gets the answer correct (Achaemenid Persia). Model gradients show that the token “subordinate ..” is impactful in the selection of an answer to the question “Macedonia was under the rule of which country?". This makes sense .. good for BERTbase. more
Jun 16, 2020 · notebook

Evaluating QA: Metrics, Predictions, and the Null Response →

by Melanie · A deep dive into computing QA predictions and when to tell BERT to zip it! In our last post, Building a QA System with BERT on Wikipedia, we used the HuggingFace framework to train BERT on the SQuAD2.0 dataset and built a simple QA system on top of the Wikipedia search engine. This time, we’ll look at how to assess the quality of a BERT-like model for Question Answering.
May 19, 2020 · notebook

Building a QA System with BERT on Wikipedia →

by Melanie · So you’ve decided to build a QA system. You want to start with something simple and general so you plan to make it open domain using Wikipedia as a corpus for answering questions. You want to use the best NLP that your compute resources allow (you’re lucky enough to have access to a GPU) so you’re going to focus on the big, flashy Transformer models that are all the rage these days.
Apr 28, 2020 · notebook

Intro to Automated Question Answering →

by Melanie · Welcome to the first edition of the Cloudera Fast Forward blog on Natural Language Processing for Question Answering! Throughout this series, we’ll build a Question Answering (QA) system with off-the-shelf algorithms and libraries and blog about our process and what we find along the way. We hope to wind up with a beginning-to-end documentary that provides:
Apr 1, 2020 · newsletter

Enterprise Grade ML

by Shioulin · At Cloudera Fast Forward, one of the mechanisms we use to tightly couple machine learning research with application is through application development projects for both internal and external clients. The problems we tackle in these projects are wide ranging and cut across various industries; the end goal is a production system that translates data into business impact. What is Enterprise Grade Machine Learning? Enterprise grade ML, a term mentioned in a paper put forth by Microsoft, refers to ML applications where there is a high level of scrutiny for data handling, model fairness, user privacy, and debuggability. more

Popular posts

Oct 30, 2019 · newsletter
Exciting Applications of Graph Neural Networks
Nov 14, 2018 · post
Federated learning: distributed machine learning with data locality and privacy
Apr 10, 2018 · post
PyTorch for Recommenders 101
Oct 4, 2017 · post
First Look: Using Three.js for 2D Data Visualization
Aug 22, 2016 · whitepaper
Under the Hood of the Variational Autoencoder (in Prose and Code)
Feb 24, 2016 · post
"Hello world" in Keras (or, Scikit-learn versus Keras)


In-depth guides to specific machine learning capabilities


Machine learning prototypes and interactive notebooks


A usable library for question answering on large datasets.

Explain BERT for Question Answering Models

Tensorflow 2.0 notebook to explain and visualize a HuggingFace BERT for Question Answering model.

NLP for Question Answering

Ongoing posts and code documenting the process of building a question answering model.

Interpretability Revisited: SHAP and LIME

Explore how to use LIME and SHAP for interpretability.


Cloudera Fast Forward is an applied machine learning reseach group.
Cloudera   Blog   Twitter