HOME ABOUT PORTFOLIO CONTACT
Juan Camilo Pérez M.Sc.

ABOUT ME

I am a Biomedical Engineer from Colombia. I studied at Universidad de los Andes in Bogotá, and obtained my Bachelor's on 2017. During my last year of undergrad I took a course on Computer Vision and I fell in love with the thing. Thus, I continued my studies by doing a Master's under Prof. Pablo Arbeláez. During my Master's I worked on Computer Vision and Natural Language Processing. Currently, I'm doing a joint Ph.D. program in Computer Vision between Universidad de los Andes, in Colombia, and KAUST, in Saudi Arabia. I'm advised by Professors Pablo Arbeláez and Bernard Ghanem.

My research interests revolve around robustness in neural networks. In particular, my current focus is in how egocentric vision--vision considered from the biased perspective of humans-- inherently induces robustness. I think there is a lot we can harness from considering how humans exploit data in (relatively) efficient ways to learn accurate and robust representations that allow us to reason in terms of abstracts concepts.

Besides research, I like reading about psychology and economics. I think both of these subjects make very interesting questions. Specifically, I think they can provide insights about how and why humans do what we do. My favorite books are "Learning How To Learn", by Oakley and Sejnowski, "Thinking, Fast and Slow" by Kahneman, "Man's Search for Meaning" by Frankl, "The Righteous Mind" by Haidt, and "Freakonomics" by Levitt and Dubner

Juan Camilo


Photo of me

ABOUT MY WORK

After graduating from my Bachelor's, I got into a M.Sc. again at the Biomedical Engineering department at Uniandes, where I was a Research Assistant.

I financed my Master's degree through a project with Colciencias, which is the colombian institution that promotes science in the country. I worked with Prof. Antonio Salazar in developing technological solutions that can improve the treatment of strokes and reduce their sequelae. The first part of this project involved the visualization of the data, regarding strokes, that has been compiled by the Ministerio de Salud y Protección Social (Ministry of Health and Social Protection).

PORTFOLIO

MY WORK

For a complete list of my publications, please visit my Google Scholar page.

Dynamic Multimodal Instance Segmentation guided by natural language queries

Here is a sample of an interesting project with Édgar Margffoy-Tuay and Emilio Botero.
We were trying to segment particular instances of objects in images based on a (natural) language query using a mixture of convolutional and recurrent neural networks.


The idea is that you provide the program an image and a short expression making reference to something in the image, and the program tries to segment the object you're referring to.
For example, here the query was "guy on the right". The output of the network should be a binary mask, having some color inside the object being referenced, and another color in the background. Nevertheless, this is not exactly the output of the network, but rather a more continuous function - some sort of mask, having high values where the network is quite sure about the presence of the object and low values everywhere else. Therefore, we need to apply some threshold at the end to produce a more detailed segmentation.


This work was published at the European Conference on Computer Vision, in 2018 (link).

Use the slider below to change the threshold of the mask that was produced by our neural network

Threshold: (slide!)

Input for the network
Un-thresholded output of the network

Visualizing data from the Ministry of Health and Social Protection

I developed a small website for visualizing data regarding strokes in Colombia. Although it is not exactly my field, I enjoyed learning about tools like HTML, CSS, JS, and some of JS' very powerful libraries like D3, DC, Crossfilter and jQuery.

Logo of the website I've been working on

Please click here to see the data visualization project.

Gabor Layers Enhance Network Robustness

We studied the way in which introducing an inductive bias in the structure of the filters learnt by Deep Neural Networks could have effects in the robustness against adversarial attacks of these networks. I think this is a very interesting avenue for future research. This work was published in the European Conference on Computer Vision in 2020 (link).

Set of filters learnt in the Gabor layer.

CONTACT

WHERE I WORK

I'd love your feedback!

KAUST, Thuwal, KSA. Building 1, Workstation 09
Email: juan.perezsantamaria AT kaust.edu.sa