From Efficient 3D Learning to 3D Foundation Models

Future AI systems are envisioned to efficiently perceive and interact with the physical (3D) world while also being able to seamlessly re-create simulated digital environments for real-time and immersive applications. My research bridges the gap between the physical and digital worlds by enabling machines to accurately perceive and reconstruct 3D geometry from visual sensors. Moreover, my work focuses on the design of a comprehensive 3D foundation model that unifies a wide range of 3D problems within a single framework and can be executed differentiably starting from unposed images, thereby reducing incremental error and engineering efforts. Building on my previous research experiences, my ongoing work examines the self-supervised pre-training of these 3D foundation models with large-scale web video data, mitigating the limitations in generalization capacity caused by the scarcity of annotated 3D data. My holistic research vision focuses on integrating geometric priors into autonomous robotics and generative AI techniques that adhere to the physical laws governing object motion and interactions in 3D space over time, thereby developing more advanced spatial intelligence.

Learning to See the World in 3D

Humans can effortlessly construct rich mental representations of the 3D world from sparse input, such as a single image. This is a core aspect of intelligence that helps us understand and interact with our surroundings and with each other. My research aims to build similar computational models–artificial intelligence methods that can perceive properties of the 3D structured world from images and videos. Despite remarkable progress in 2D computer vision, 3D perception remains an open problem due to some unique challenges, such as limited 3D training data and uncertainties in reconstruction. In this talk, I will discuss these challenges and explain how my research addresses them by posing vision as an inverse problem, and by designing machine learning models with physics-inspired inductive biases. I will demonstrate techniques for reconstructing 3D faces and objects, and for reasoning about uncertainties in scene reconstruction using generative models. I will then discuss how these efforts advance us toward scalable and generalizable visual perception and how they advance application domains such as robotics and computer graphics.

Learning under Inexact Data and Model for Computational Imaging

Computational imaging is an indispensable tool for recovering images from their indirect and noisy measurements through the application of advanced algorithms. Its scope ranges from enhancing the photos taken by a camera to providing high-fidelity medical images for diagnosis. The recent years have witnessed remarkable progress in machine learning (ML) techniques, which leads to a transformative paradigm shift in computational imaging algorithms. This new paradigm seeks to integrate the physics of the imaging systems and the latest ML techniques. While such learning-based approaches have demonstrated the state-of-the-art performance, they come with several key practical challenges that must be addressed before we can deploy them to the real-world imaging systems. This talk focuses on two specific challenges: (1) the challenge of learning under inexact data, where the data useful for learning is scarce (e.g., the scarcity of learning reference in medical images), and (2) the challenge of learning under an inexact model, where the forward model characterizing the underlying physics of the imaging system is imprecise. This presentation introduces new algorithm designs to address these challenges with theoretical analysis to provide mathematical guarantees. The applicability and effectiveness of these algorithms has been substantiated through empirical validation on a range of real-world biomedical imaging systems.

Making 2D Diffusion Models 3D-Aware

Diffusion-based image generators, such as Stable Diffusion and DALL-E, enable creating 2D images from text descriptions. In this talk, we will cover recent research towards making these models 3D-aware, allowing them to generate consistent images of a 3D scene. It will cover approaches to incorporate camera control into diffusion models and different techniques for explicitly or implicitly representing the 3D scene. Furthermore, we will discuss the role of high-quality, large-scale, synthetic and real datasets in this task. Finally, we will conclude with a discussion of future research directions in this area and key takeaways from past work.

Computational imaging in the Era of Large-Scale Foundation Models

Foundation models leverage the strength of self-supervised learning on large data volumes to learn complex real-world patterns. These large models have been found to have multi-task generalization capabilities and are widely regarded as the future of computer vision and AI. A pertinent question is whether scale is the solution to all problems, in which case the benefits of computational imaging are largely rendered ineffective. We argue that this is not the case. We begin by discussing classes of problems where unbounded scale is a fundamentally incapable solution, using the case study of equitable imaging. We will then discuss classes of problems where unbounded scale is not practical, using the case study of medical imaging. Finally, we will discuss the interplay between computational imaging and large-scale foundation models, where foundation models and their underlying diverse priors can aid sensor design, while novel sensor data can improve the capabilities of foundation models. With this context, we will discuss the future of computational imaging, in the era of large foundation models.

Adapting and Regularizing Diffusion Models for Inverse Problems

Diffusion models are revolutionizing the field of inverse imaging by leveraging powerful foundational generative priors. This talk will initially explore methods for adapting diffusion models to inverse problems under mismatched priors, specifically out-of-distribution (OOD) measurements, even with only a single measurement. This significantly broadens the practical utility of diffusion models in real-world scenarios where acquiring high-quality training data is challenging. Subsequently, the discussion will shift to the application of a similar approach in resolving ambiguous elements, such as text conditioning, in foundational diffusion models. Additionally, an efficient method for regularizing the space of the posterior distribution with text conditions will be introduced. Although these applications may appear distinct, they will be unified within a comprehensive framework, highlighting new possibilities for solving inverse problems in imaging.

Extreme Novel-view Synthesis and 3D Content Creation Using Deep Generative Prior

Virtual reality and augmented reality bring increasing demand for 3D content generation. However, the creation of such content often involves laborious tasks performed by human experts. I will discuss our recent study on the novel-view synthesis and 3D content generation in a more challenging context. Specifically, we explore the task of transforming a single image into a 3D object and demonstrate, for the first time, the capability to generate a realistic 3D object with 360-degree views that closely align with the provided reference image. Our proposed framework, NeuralLift-360, leverages large-scale text-to-image generative models, such as the Diffusion Model, to fulfill the everlasting curiosity of synthesizing novel views of objects from a single image. Extensive experiments demonstrate that NeuralLift-360 significantly outperforms existing state-of-the-art baselines.

Generative Models and Computational Imaging: Soft Diffusion and Learning from Corrupted Data

Diffusion models have revolutionized the field of generative modeling, giving rise to systems that generate images of unparalleled quality. In this talk, we will talk about recent algorithmic innovations that generalize diffusion models and tackle crucial limitations of the standard formulation. In the first part of the presentation, we will introduce a broader family of corruption processes and demonstrate how to reverse them using a new framework, Soft Diffusion, that delivers state-of-the-art results. Following this, we will discuss ongoing research efforts to train diffusion models to learn distributions from corrupted samples. Throughout the talk, we will reveal connections between diffusion models and the realm of inverse problems, highlighting possible avenues for further research in advancing generative models.

Beyond RGB: Next-Gen Cameras for Untangling Light Transport

Images result from a complex interplay between light and objects in a scene. Capturing and decoding this interplay, known as light transport, is crucial for generating photorealistic digital twins, uncovering hidden information, and enhancing scene understanding from images. However, traditional RGB cameras are ill-suited for inferring light transport due to inherent ambiguities. In this talk, I will discuss how next-generation cameras can disentangle light transport by capturing unique dimensions of light, such as time-of-flight and polarization. I will explain how polarization is a useful and practical cue for inverse rendering. I will show how snapshot polarization cameras enable reflectance separation of dynamic scenes and radiance decomposition under unstructured lighting. I will also describe how measuring the time-of-flight distribution of light allows us to see around the corner and how the advances in conventional line-of-sight imaging can inspire us to design effective non-line-of-sight imaging systems.

Uncertain Surface Reconstruction

We propose a method to introduce uncertainty to the surface reconstruction problem. Specifically, we introduce a statistical extension of the classic Poisson Surface Reconstruction algorithm for recovering shapes from 3D point clouds. Instead of outputting an implicit function, we represent the reconstructed shape as a modified Gaussian Process, which allows us to conduct statistical queries (e.g., the likelihood of a point in space being on the surface or inside a solid). We show that this perspective improves PSR's integration into the online scanning process, broadens its application realm, and opens the door to other lines of research such as applying task-specific priors.

Seeing Beneath the Skin with Computational Photography

For more than two decades, telehealth held the promise of greater access to healthcare services in the home but remained as a niche opportunity due to a combination of regulatory, economic, and cultural barriers that prevented the expansion and innovation of digital health services. The SARS-CoV-2 pandemic promoted a rapid and explosive growth of online services that provided an effective mechanism for safely providing care without risk of exposure. Remote patient monitoring is now gaining adoption and new physiological and medical imaging modalities that leverage recent advances in computational photography are emerging. These methods use everyday sensors to non-invasively inspect and measure the internal state of the body. I will present research on physiological and behavioral measurement via ubiquitous hardware, and highlight approaches that capture multimodal signals (e.g., facial and motor movements, heart rate and HRV, respiration, blood pressure) without contact with the body. I'll show examples of state-of-the-art, on-device neural models and a synthetics data pipeline to help us learn more robust representations and achieve performance close to that of contact sensors. Following this, I will give examples of novel human-computer interfaces that leverage these signals to improve health, wellbeing and communication.

Computational Imaging of the Black Hole at the Center of our Galaxy

The Event Horizon Telescope (EHT) is a unique computational camera with a goal of imaging the glowing fluid surrounding supermassive black holes. In May of 2022, the EHT collaboration revealed the first images of the black hole at the center of our galaxy: Sagittarius A* (Sgr A*). These images were computationally reconstructed from measurements taken by synchronized telescopes around the globe. Imaging Sgr A* was challenging due to its evolution, which is dynamic on the timescale of acquisition. The test our imaging pipelines in a challenging, realistic setting we developed stochastic dynamic models that mimic EHT observations of Sgr A*. While images certainly offer interesting insights, looking toward the future, we are developing new computational algorithms that aim to go beyond a 2D image. For example, could we use EHT observations to recover the dynamic evolution or even the 3D structure? In this talk, I will highlight new and exciting prospects for the future of black-hole imaging. By peeling away different layers of the underlying physics, we show how computational algorithms could shed light on complex dynamic processes. Our hope is that in the not-too-distant future, they will enable scientific discovery and even provide a glimpse into the very nature of space-time itself in our galaxy's most extreme environment.

Computational Imaging: Chloroplasts to Tree Canopies

Computational imaging is both a highly abstract and extremely detailed field. From 40,000 feet, we formulate how sensing hardware and imaging software can work together optimally. But if we zoom in, each application space and sensing task will demand unique computational imaging solutions. In this talk, I’ll discuss two computational imaging problems I’ve encountered in two very different domains. First, I’ll talk about 3D quantitative refractive index imaging of microscopic phase objects. I will introduce a novel microscope design with increased measurement diversity to improve the 3D imaging system. Second, I’ll discuss remote sensing imaging spectroscopy, or hyperspectral imaging, for Earth science applications. I’ll show how leveraging atmospheric smoothness can enable us to better image surface properties of the Earth. I will then discuss the commonalities I’ve seen in computational imaging across scales and applications, and how we might improve our computational imaging frameworks for chloroplasts to tree canopies and everything in between.

Rethinking Supervision for Some Vision Tasks

Deep neural networks are highly parameterized functions that have seen empirical successes across a number of computer vision tasks. Due to their size, they require tens of thousands to millions of training examples, which typically come in the form of an image or images and human annotated ground truth. Curating vision datasets, in general, amounts to numerous man-hours; tasks like depth estimation requires an even more massive effort. In this talk, I will introduce alternative forms of supervision that leverage geometric constraints, synthetic data, or the abundance of pretrained models available online for depth estimation, which has traditionally relied largely on expensive manual labeling. I will show that doing so can lead to smaller and much faster models that are comparable in performance to state-of-the-art methods trained on human annotated ground truth. I will also discuss single image deraining, which unlike most vision tasks, have traditionally relied on synthetic data for supervision – this is largely due to the intractability of collecting the ideal image and ground truth pair for learning to remove rain. I will introduce another notion of “ground truth”, which enables us to collect a large scale dataset of real data, that does not suffer from the sim2real gap, and conclude with an invitation to participate in a sponsored challenge centered around the dataset.

Imaging, Fast and Slow: Computational Imaging for Sensing High-speed Phenomena

The world is full of phenomena that are either too fast or too minute for our eyes to observe. Moreover, despite recent advances in sensor technologies, most cameras are just as blind to these phenomena as our eyes. One reason for this limitation is that capturing high-speed video at high-spatial resolutions is fundamentally challenging. Computational imaging is an exciting field that allows us to build novel vision systems that can capture these high-speed phenomena, consequently revealing additional hidden information about our environments. In this talk, I will cover three projects using computational imaging to sense various fast phenomena. First, I will describe the ACam - a camera designed to capture the minute flicker of artificial lights ubiquitous in our modern environments. I will show that bulb flicker is a powerful visual cue that enables various applications like scene light source unmixing, reflection separation, and remote analyses of the electric grid itself. Second, I will describe Diffraction Line Imaging, a novel imaging principle that exploits diffractive optics to capture sparse 2D scenes with 1D (line) sensors. The method's applications include fast motion capture and structured light 3D scanning with line illumination and line sensing. Lastly, I will present a novel method for sensing minute high-frequency surface vibrations (up to 63kHz) for multiple scene sources simultaneously, using "slow" sensors rated for only 130Hz operation. Applications include capturing vibration caused by audio sources (e.g., speakers, human voice, and musical instruments) and analyzing the vibration modes of objects.

Shady business: Occluder-aided Non-Line-of-Sight Imaging

Non-line-of-sight (NLOS) imaging involves the reconstruction of scenes that are not directly visible to the observer. Typically, this is based upon measurements of light originating or returning from the scene that has then scattered from a diffuse reflector (such as a wall) towards the observer. The results of NLOS imaging range from the recovery of 2D photograph-like images of a hidden area, to determining the position, motion or number of hidden objects, to 3D reconstructions of a hidden space. In this talk, I will discuss two rather different NLOS imaging methods that we have developed: Computational Periscopy, and Edge-Resolved Transient Imaging. Although the methods differ greatly in their aims and mechanisms, they still share a common thread: the opportunistic use of occluders (objects that block light) that are ubiquitous to the spaces around us.

Computational Interferometry

Optical interferometry is a technique that uses the wave-optics phenomenon of interference caused by the superposition of two or more light waves to measure the properties of the waves or of the scenes they interact with. Due to its capability to resolve micron-scale displacements, interferometry has found applications in physics, biomedical imaging, and precision fabrication. We introduce a class of new interferometric techniques, collectively called computational interferometry, that allows combining these capabilities with precise, micron-scale control on the set of light paths imaged by the camera in the scene of interest. We show how to design and build prototype optical systems to implement computational interferometry specialized to a few application settings: imaging through scattering and micron-scale 3D shape acquisition. We further show how computational interferometry allows us to perform these tasks robustly to adverse environmental conditions such as global illumination, ambient lighting, and environment mechanical noise. Lastly, we showcase an interferometric system that enables passive time-of-flight imaging using only sunlight as a light source, with micron-scale depth reconstructions captured in uncontrolled outdoor environments with an interferometer built on an ordinary utility cart. Our computational interferometry techniques find applications in seeing through scattering media such as tissue, inspection, and precision manufacturing.

Bio-inspired depth sensing using computational optics

Computational visual sensing combines efficient computation and novel optics to provide computer vision systems that minimize size, weight, and power while maximizing speed, accuracy, and functionality. In these systems, signals are processed both optically and electronically, using the strengths of each to maximally exploit the visual characteristics of an environment. Computational visual sensors are often inspired by nature, where invertebrates and other small animals have evolved with optics and neural wirings that synergize to perceive with extremely small size and power. In this talk, I will present a class of small computational sensors for passively measuring depth. They are inspired by the eyes of jumping spiders, which have specialized, multi-layered retinae and sense depth using optical defocus. We make three main contributions in inventing these sensors. First, we propose a novel, physics-based depth cue leveraging differential changes of image defocus. The depth cue takes the form of a simple analytical expression and is robust to optical non-idealities. Second, we incorporate the physics-based depth cue into the design of a neural network, which yields an efficient computational graph that predicts per-pixel depth and confidence at below 700 floating-point operations (FLOPs) per output pixel. Third, we designed two optical setups to pair with the computation. One setup consists of a deformable lens and has an extended working range due to its ability to perform optical accommodation. The other uses a specially-designed metasurface, which is an ultra-thin, flat nanophotonic device with versatile wavefront shaping capability. It is one of the world’s first nanophotonic depth sensors. Both setups are monocular and compact, and both can measure scene depth at 100 frames per second. Our computational depth sensors are one example of how optics and computation will evolve in the future to achieve artificial visual sensing on tiny platforms where vision is currently impossible.

Video denoising in starlight using a learned, physics-informed noise model

Some animals, such as hawkmoths and carpenter bees, can effectively navigate on the darkest moonless nights by the light of the stars (< 0.001 lux), whereas our best CMOS cameras generally require at least 3/4 moon illumination (> 0.1 lux) to image moving objects at night. Videography on moonless, clear nights is extremely challenging due to low photon counts. Long exposures can be used, but precludes the imaging of moving objects. Alternatively, higher gain can make pixels more sensitive to light, but leads to increased noise which is often non-Gaussian, sensor-specific, and difficult to model or characterize. Off-the-shelf denoisers which assume Gaussian noise often fail in these low light, high gain regimes due to this structured noise. In this work, we present a learned, physics-informed noise model which, after training, can more accurately synthesize camera noise at the lowest light and highest gain settings. Using our noise model, we train a video denoiser using a combination of synthetic noisy video clips and real still images. We capture a 5-10 fps video dataset with significant motion on a moonless, clean night with no active illumination (0.6-0.7 mililux). Using our video denoiser, we demonstrate photorealistic video denoising in starlight (< 0.001 lux) for the first time.

Computational imaging with multiply scattered photons

Computational imaging has advanced to a point where the next significant milestone is to image in the presence of multiply-scattered light. Though traditionally treated as noise, multiply-scattered light carries information that can enable previously impossible imaging capabilities, such as imaging around corners and deep inside tissue. The combinatorial complexity of multiply-scattered light transport makes it necessary to use increasingly complex imaging systems to make imaging with multiply-scattered light feasible; examples include time-of-flight, structured-light, and acousto-optic imaging systems. The combined complexity of physics and systems makes the optimization of imaging of multiply-scattered light a challenging, high-dimensional design problem. In my research, I utilize graphics and physics-based rendering to explore this complex design space and create imaging systems that optimally sense multiply-scattered light. I will show two examples of this approach. First, I will discuss how to develop rendering tools for time-of-flight cameras, and how to use these tools to design and build optimal time-of-flight systems for non-line-of-sight imaging. Second, I will discuss how to simulate continuously-refractive radiative transfer and use such simulations to optimize acousto-optic systems for imaging inside tissue and other scattering media.

Utilizing Nature’s Limits for better Computational 3D Imaging

Computational 3D imaging and 3D display principles are ‘enabling technologies’ that have the potential to foster transformational technical changes in the next decades. The list of possible future applications scenarios is long: Novel breeds of cameras could see through tissue, fog, or around corners. Portable medical 3D imaging and 3D display devices could effectively assist doctors in diagnosis and therapy. Precise and fast 3D scanners could become essential for measuring and analyzing dynamic scenes during robotic surgery, industrial inspection, autonomous navigation, or cultural heritage applications. Novel 3D display and eye-tracking methods could finally enable immersive AR/VR experiences. These are just a few examples. In midst of these seemingly endless possibilities, the knowledge about fundamental physical and information-theoretical limits in imaging proves to be a powerful tool: By knowing that our imaging device already operates at the physical limit (e.g., of resolution), we can avoid unnecessary effort and investments in better hardware such as faster detectors, or cameras with higher pixel resolution. Moreover, limits often appear as uncertainty products, making it possible to optimize our system towards a specific quantity (e.g., speed) by trading in information less critical for the respective application. Although the imaging device is essential in this optimization, the central role is assumed by the illumination, which serves as encoder of the desired information. In this talk, I will discuss the virtue of limits and merit of illumination modalities in computational 3D imaging systems using the examples of my research discussed above. Among other systems, I will introduce a novel method to image hidden objects through scattering media and around corners, and a ‘single-shot 3D camera’ for the motion-robust, precise, and dense 3D measurement of fast live scenes.

Foundational Geometric Vision and its Role in Modern 3D Data-Acquisition Methods

In the coming decade, dense, detailed, and precise 3D data acquisition of objects or scenes is sure to become one of the most important problems in computer vision and industrial machine vision. Moreover, it can be helpful for a wide range of other cutting-edge scientific disciplines such as metrology, geometry processing, forensics, etc. Unfortunately, at present, we don't have a general-purpose vision algorithm or framework for dense 3D data acquisition to meet the required precision. In this talk, I will present a few of our recent ideas and algorithms showing how a mindful use of existing classical geometric ideas can help us get closer to the goal of high-fidelity 3D data acquisition from images or available depth-sensing modalities. Furthermore, the talk will briefly discuss the dense 3D data acquisition methods using an online robot system and its current challenges. To conclude, this talk will pivot around the blueprint of classical geometric vision ideas and how to use it to effectively exploit the strength of deep neural networks for more trustworthy 3D data acquisition.

Artificial Intelligence in Dermatology: Opportunities and Challenges

The clinical practice of dermatology is visual in nature: physicians make decisions after either directly assessing the skin or viewing clinical photos. Deep learning algorithms have the potential to streamline processes in dermatology and teledermatology. However, the development of generalizable and robust clinical algorithms will require diverse, high quality datasets. I will discuss our work in developing a deep learning algorithm for assessing photo quality in teledermatology. After completing initial in silico testing, we are translating this work from “bytes” to “bedside”. Additionally, I will discuss the challenges for developing fair and robust algorithms in dermatology and what steps we can take to address these challenges.

Computational astronomical imaging: inference and sensing algorithms for future black hole observations

Recent development in computational imaging has led to a series of great discoveries in astronomy, such as the first image of a black hole (M87*) acquired by the Event Horizon Telescope (EHT). In this talk, I will introduce two new computational imaging algorithms we have developed for advancing future black hole observations. First, I will introduce Deep Probabilistic Imaging: a novel variational inference algorithm that trains a deep generative model to quantify the uncertainty of a reconstructed image. This approach helps to better characterize the confidence of structures in a black hole reconstruction, leading to more reliable scientific interpretations. Second, I will present “cosense”, a physics-informed deep learning approach that jointly optimizes a telescope network design with an image reconstruction approach. This algorithm is currently being used to select new telescope locations for the next generation Event Horizon Telescope (ngEHT). I will also briefly show the general applications of these algorithms in other scientific imaging tasks, such as exoplanet discovery and MRI.

One Photon at a Time: Compensating for Nonideal Electronics in LiDAR Imaging

Forming a digital image—with a conventional camera or computational imaging system—requires converting properties of light into a sequence of bits. The mechanisms transforming optical energy into digital signals can often be idealized or ignored, but problems such as quantization or saturation become noticeable when imaging at the limits of the sensors. One such example arises in the case of single-photon lidar, which aims to form 3D images from individual photon detections. In this talk, we consider two factors that prevent perfect recording of each photon arrival time with infinite precision: finite temporal quantization and missed detections during detector dead times. We show that incorporating nonidealities into our acquisition models significantly mitigate their effect on our ability to form accurate depth images.

Enabling an Image-Based Graphics Pipeline with Neural Radiance Fields

Neural Radiance Fields (NeRFs) have recently emerged as an effective and simple solution for recovering 3D representations of complex objects and scenes from captured images. However, there is still a long way to go before we will be able to use NeRF-like representations of real-world content in graphics pipelines as easily as standard computer graphics representations of artist-designed content, such as textured triangle meshes. I will discuss and review NeRF, and then talk about some of our recent work towards extending NeRF to enable more of the functionality we expect from the 3D representations we use in computer graphics.

Digital Health and Wellbeing: Data-Driven and Human-Centered Personalized and Adaptive Assistant

Imagine 24/7 rich human multimodal data could identify changes in physiology and behavior, and provide personalized early warnings to help you, patients, or clinicians for making better decisions or behavioral changes to support health and wellbeing. I will introduce a series of studies, algorithms, and systems we have developed for measuring, predicting, and supporting personalized health and wellbeing for clinical populations as well as people at increased risk of adverse events, including ongoing COVID-19 related projects. I will also discuss challenges, learned lessons, and potential future directions in digital health and wellbeing research.

Differential Defocus in Cameras and Microscopes

Image defocus provides a useful depth cue in computer vision, and can also be used to recover phase information in coherent microscopy. In a differential setting, both problems can be addressed by solving simple equations, known as Depth from Differential Defocus and the Transport of Intensity Equation. Relating these governing equations requires putting them on equal footing, so we'll look at the assumptions common to photography and microscopy applications, and go through a gentle introduction to coherence, light fields and Wigner Distribution Functions, and generalized phase. We'll show that depth from defocus can be seen as a special case of phase recovery, with a new interpretation of phase for incoherent settings.

Can cameras really measure vital signs? Algorithms and systems for camera-based health monitoring in unconstrained settings.

Imagine when you looked at someone, you could see their heartbeat. A suite of techniques called imaging photoplethysmography has recently enabled contactless measurements of vital signs with cameras by leveraging small intensity changes in the skin caused by cardiac activity. Measuring vital signs remotely is advantageous in several applications, including virtual doctor appointments, especially relevant during a pandemic, as well as more comfortable sleep monitoring, or monitoring of prematurely born infants. However, the camera-based physiological signals are very weak and easily corrupted by varying illumination, video compression artifacts, and head motion. Therefore, most existing methods only work in controlled settings and fail in realistic applications. We developed a denoising deep learning algorithm based on convolutional attention networks that can faithfully recover physiological signals even from heavily corrupted videos. Moreover, our denoising algorithm can recover subtle waveform dynamics, previously not possible to measure with cameras. We also discuss how we can improve the performance of deep learning methods and avoid overfitting when training on very small and not diverse datasets.