VASC Seminar
Detecting Image Synthesis — Shallow and Deep
Abstract: The proliferation of synthetic media are subject to malicious usages such as disinformation campaigns, posing potential threats to media integrity and democracy. A way to combat this is developing forensics algorithms to identify manipulated media. In the beginning of the talk, I will discuss how one can train a model to detect photos manipulated [...]
Deep Learning to Distinguish Recalled but Benign Mammography Images in Breast Cancer Screening
Abstract: Breast cancer screening using the standard mammography exam currently exhibits a high false recall rate (11.6% for women in the U.S.). Only a low proportion (0.5%) of women who were recalled for additional workup were actually found to have breast cancer. As a result of the unnecessary stress and follow-up work from these false [...]
The Plenoptic Camera
Abstract: Imagine a futuristic version of Google Street View that could dial up any possible place in the world, at any possible time. Effectively, such a service would be a recording of the plenoptic function—the hypothetical function described by Adelson and Bergen that captures all light rays passing through space at all times. While the plenoptic function [...]
Photorealistic Reconstruction of Landmarks and People using Implicit Scene Representation
Abstract: Reconstructing scenes to synthesize novel views is a long standing problem in Computer Vision and Graphics. Recently, implicit scene representations have shown novel view synthesis results of unprecedented quality, like the ones of Neural Radiance Fields (NeRF), which use the weights of a multi-layer perceptron to model the volumetric density and color of a [...]
Towards Discriminative and Domain-Invariant Feature Learning
Abstract: Deep neural networks have achieved great success in various visual applications, when trained with large amounts of labeled in-domain data. However, the networks usually suffer from a heavy performance drop on the data whose distribution is quite different from the training one. Domain adaptation methods aim to deal with such performance gap caused by [...]
Learning Efficient Visual Representation on Model, Data, Label and Beyond
Abstract: Efficient deep learning is a broad concept that we aim to learn compressed deep models and develop training algorithms to improve the efficiency of model representations, data and label utilization, etc. In recent years, deep neural networks have been recognized as one of the most effective techniques for many learning tasks, also, in the [...]
Self-supervised Learning and Generalization
Abstract: Contrastive self-supervised learning is a highly effective way of learning representations that are useful for, i.e. generalise, to a wide range of downstream vision tasks and datasets. In the first part of the talk, I will present MoCHi, our recently published contrastive self-supervised learning approach (NeurIPS 2020) that is able to learn transferable representations [...]
Learning to see from few labels
Abstract: Computer vision systems today exhibit a rich and accurate understanding of the visual world, but increasingly rely on learning on large labeled datasets to do so. This reliance on large labeled datasets is a problem especially when one considers difficult perception tasks, or novel domains where annotations might require effort or expertise. We thus [...]
Seeing the unseen: inferring unobserved information from multi-modal data
Abstract: As humans we can never fully observe the world around us and yet we are able to build remarkably useful models of it from our limited sensory data. Machine learning problems are often required to operate in a similar setup, that is the one of inferring unobserved information from the observed one. Partial observations [...]
Towards AI for 3D Content Creation
Abstract: 3D content is key in several domains such as architecture, film, gaming, and robotics. However, creating 3D content can be very time consuming -- the artists need to sculpt high quality 3d assets, compose them into large worlds, and bring these worlds to life by writing behaviour models that "drives" the characters around in [...]
Understanding the Placenta: Towards an Objective Pregnancy Screening
Abstract: My research focusses on the development of a pregnancy screening tool, that will be: (i) system and user-independent; and (ii) provides a quantifi able measure of placental health. With this end, I am working towards the design of a multiparametric quantitative ultrasound (QUS) based placental tissue characterization method. The method would potentially identify the [...]
Relational Reasoning for Multi-Agent Systems
Abstract: Multi-agent interacting systems are prevalent in the world, from purely physical systems to complicated social dynamics systems. The interactions between entities / components can give rise to very complex behavior patterns at the level of both individuals and the whole system. In many real-world multi-agent interacting systems (e.g., traffic participants, mobile robots, sports players), [...]
Self-supervised learning for visual recognition
Abstract: We are interested in learning visual representations that are discriminative for semantic image understanding tasks such as object classification, detection, and segmentation in images/videos. A common approach to obtain such features is to use supervised learning. However, this requires manual annotation of images, which is costly, ambiguous, and prone to errors. In contrast, self-supervised [...]
Reasoning over Text in Images for VQA and Captioning
Abstract: Text in images carries essential information for multimodal reasoning, such as VQA or image captioning. To enable machines to perceive and understand scene text and reason jointly with other modalities, 1) we collect the TextCaps dataset, which requires models to read and reason over text and visual content in the image to generate image [...]
Point Cloud Registration with or without Learning
Abstract: I will be presenting two of our recent works on 3D point cloud registration: A scene flow method for non-rigid registration: I will discuss our current method to recover scene flow from point clouds. Scene flow is the three-dimensional (3D) motion field of a scene, and it provides information about the spatial arrangement [...]
Propelling Robot Manipulation of Unknown Objects using Learned Object Centric Models
Abstract: There is a growing interest in using data-driven methods to scale up manipulation capabilities of robots for handling a large variety of objects. Many of these methods are oblivious to the notion of objects and they learn monolithic policies from the whole scene in image space. As a result, they don’t generalize well to [...]
When and Why Does Contrastive Learning Work?
Abstract: Contrastive learning organizes data by pulling together related items and pushing apart everything else. These methods have become very popular but it's still not entirely clear when and why they work. I will share two ideas from our recent work. First, I will argue that contrastive learning is really about learning to forget. Different [...]
Anticipating the Future: forecasting the dynamics in multiple levels of abstraction
Abstract: A key navigational capability for autonomous agents is to predict the future locations, actions, and behaviors of other agents in the environment. This is particularly crucial for safety in the realm of autonomous vehicles and robots. However, many current approaches to navigation and control assume perfect perception and knowledge of the environment, even though [...]
Learning to Perceive Videos for Embodiment
Abstract: Video understanding has achieved tremendous success in computer vision tasks, such as action recognition, visual tracking, and visual representation learning. Recently, this success has gradually been converted into facilitating robots and embodied agents to interact with the environments. In this talk, I am going to introduce our recent efforts on extracting self-supervisory signals and [...]
Open Challenges in Sign Language Translation & Production
Abstract: Machine translation and computer vision have greatly benefited of the advances in deep learning. The large and diverse amount of textual and visual data have been used to train neural networks whether in a supervised or self-supervised manner. Nevertheless, the convergence of the two field in sign language translation and production is still poses [...]
3D Recognition with self-supervised learning and generic architectures
Abstract: Supervised learning relies on manual labeling which scales poorly with the number of tasks and data. Manual labeling is especially cumbersome for 3D recognition tasks such as detection and segmentation and thus most 3D datasets are surprisingly small compared to image or video datasets. 3D recognition methods are also fragmented based on the type [...]
Rapid Adaptation for Robot Learning
Abstract: How can we train a robot to generalize to diverse environments? This question underscores the holy grail of robot learning research because it is difficult to supervise an agent for all possible situations it can encounter in the future. We posit that the only way to guarantee such a generalization is to continually learn and [...]
Humans, hands, and horses: 3D reconstruction of articulated object categories using strong, weak, and self-supervision
Abstract: Reconstructing 3D objects from a single 2D image is a task that humans perform effortlessly, yet computer vision so far has only robustly solved 3D face reconstruction. In this talk we will see how we can extend the scope of monocular 3D reconstruction to more challenging, articulated categories such as human bodies, hands and [...]
Looking behind the Seen in Order to Anticipate
Abstract: Despite significant recent progress in computer vision and machine learning, personalized autonomous agents often still don’t participate robustly and safely across tasks in our environment. We think this is largely because they lack an ability to anticipate, which in turn is due to a missing understanding about what is happening behind the seen, i.e., [...]
The Clinician’s AI Partner: Augmenting Clinician Capabilities Across the Spectrum of Healthcare
Abstract: Clinicians often work under highly demanding conditions to deliver complex care to patients. As our aging population grows and care becomes increasingly complex, physicians and nurses are now also experiencing feelings of burnout at unprecedented levels. In this talk, I will discuss possibilities for computer vision to function as a partner to clinicians, and to augment their capabilities, across [...]
Reliable and Accessible Visual Recognition
Abstract: As visual recognition models are developed across diverse applications; we need the ability to reliably deploy our systems in a variety of environments. At the same time, visual models tend to be trained and evaluated on a static set of curated and annotated data which only represents a subset of the world. In this [...]
Fake It Till You Make It: Face analysis in the wild using synthetic data alone
Abstract: In this seminar I will demonstrate how synthetic data alone can be used to perform face-related computer vision in the wild. The community has long enjoyed the benefits of synthesizing training data with graphics, but the domain gap between real and synthetic data has remained a problem, especially for human faces. Researchers have tried [...]
Leveraging StyleGAN for Image Editing and Manipulation
Abstract: StyleGAN has recently been established as the state-of-the-art unconditional generator, synthesizing images of phenomenal realism and fidelity, particularly for human faces. With its rich semantic space, many works have attempted to understand and control StyleGAN’s latent representations with the goal of performing image manipulations. To perform manipulations on real images, however, one must learn to [...]
Next-Gen Video Communication
Abstract: Video communication connects our world. It is necessary in conducting business, educational and personal activities across different geographical locations. However, the quality of an average user’s video communication is dramatically worse than that of professionally created videos in news broadcasts, talk shows, and on YouTube. This is because professionally created videos are often captured with [...]
Activity Understanding of Scripted Performances
Abstract: The PSU Taichi for Smart Health project has been doing a deep-dive into vision-based analysis of 24-form Yang-style Taichi (TaijiQuan). A key property of Taichi, shared by martial arts katas and prearranged form exercises in other sports, is practice of a scripted routine to build both mental and physical competence. The scripted nature of routines [...]
Domain adaptive object detection
Abstract: Recent advances in deep learning have led to the development of accurate and efficient models for object detection. However, learning highly accurate models relies on the availability of large-scale annotated datasets. Due to this, model performance drops drastically when evaluated on label-scarce datasets having visually distinct images. Domain adaptation tries to mitigate this degradation. In [...]
Visual Understanding across Semantic Groups, Domains and Devices
Abstract: Deep neural networks often lack generalization capabilities to accommodate changes in the input/output domain distributions and, therefore, are inherently limited by the restricted visual and semantic information contained in the original training set. In this talk, we argue the importance of the versatility of deep neural architectures and we explore it from various perspectives. [...]
Topology-Driven Learning for Biomedical Imaging Informatics
Abstract: Thanks to decades of technology development, we are now able to visualize in high quality complex biomedical structures such as neurons, vessels, trabeculae and breast tissues. We need innovative methods to fully exploit these structures, which encode important information about underlying biological mechanisms. In this talk, we explain how topology, i.e., connected components, handles, loops, [...]
Learning generative representations for image distributions
Abstract: Autoencoder neural networks are an unsupervised technique for learning representations, which have been used effectively in many data domains. While capable of generating data, autoencoders have been inferior to other models like Generative Adversarial Networks (GAN’s) in their ability to generate image data. We will describe a general autoencoder architecture that addresses this limitation, and [...]
Building Intelligent and Visceral Machines: From Sensing to Application
Abstract: Humans have evolved to have highly adaptive behaviors that help us survive and thrive. As AI prompts a move from computing interfaces that are explicit and procedural to those that are implicit and intelligent, we are presented with extraordinary opportunities. In this talk, I will argue that understanding affective and behavioral signals presents many opportunities [...]
GANcraft – an unsupervised 3D neural method for world-to-world translation
Abstract: Advances in 2D image-to-image translation methods, such as SPADE/GauGAN, have enabled users to paint photorealistic images by drawing simple sketches similar to those created in Microsoft Paint. Despite these innovations, creating a realistic 3D scene remains a painstaking task, out of the reach of most people. It requires years of expertise, professional software, a library [...]
Learning Optical Flow: Model, Data, and Applications
Abstract: Optical flow provides important information about the dynamic world and is of fundamental importance to many tasks. In this talk, I will present my work on different aspects of learning optical flow. I will start with the background and talk about PWC-Net, a compact and effective model built using classical principles for optical flow. Next, [...]
Do Vision-Language Pretrained Models Learn Spatiotemporal Primitive Concepts?
Abstract: Vision-language models pretrained on web-scale data have revolutionized deep learning in the last few years. They have demonstrated strong transfer learning performance on a wide range of tasks, even under the "zero-shot" setup, where text "prompts" serve as a natural interface for humans to specify a task, as opposed to collecting labeled data. These models are [...]
Max-Affine Spline Insights into Deep Learning
Abstract: We build a rigorous bridge between deep networks (DNs) and approximation theory via spline functions and operators. Our key result is that a large class of DNs can be written as a composition of max-affine spline operators (MASOs) that provide a powerful portal through which we view and analyze their inner workings. For instance, [...]
Understanding 3D Scenes and Interacting Hands
Abstract: Abstract: The long-term goal of my research is to help computers understand the physical world from images, including both 3D properties and how humans or robots could interact with things. This talk will summarize two recent directions aimed at enabling this goal. I will begin with learning to reconstruct full 3D scenes, including [...]
Multimodal Modeling: Learning Beyond Visual Knowledge
Abstract: The computer vision community has embraced the success of learning specialist models by training with a fixed set of predetermined object categories, such as ImageNet or COCO. However, learning only from visual knowledge might hinder the flexibility and generality of visual models, which requires additional labeled data to specify any other visual concept and [...]
Audio-Visual Learning for Social Telepresence
Abstract Relationships between people are strongly influenced by distance. Even with today’s technology, remote communication is limited to a two-dimensional audio-visual experience and lacks the availability of a shared, three-dimensional space in which people can interact with each other over the distance. Our mission at Reality Labs Research (RLR) in Pittsburgh is to develop such [...]
Representations in Robot Manipulation: Learning to Manipulate Ropes, Fabrics, Bags, and Liquids
Abstract: The robotics community has seen significant progress in applying machine learning for robot manipulation. However, much manipulation research focuses on rigid objects instead of highly deformable objects such as ropes, fabrics, bags, and liquids, which pose challenges due to their complex configuration spaces, dynamics, and self-occlusions. To achieve greater progress in robot manipulation of [...]
Towards editable indoor lighting estimation
Abstract: Combining virtual and real visual elements into a single, realistic image requires the accurate estimation of the lighting conditions of the real scene. In recent years, several approaches of increasing complexity---ranging from simple encoder-decoder architecture to more sophisticated volumetric neural rendering---have been proposed. While the quality of automatic estimates has increased, they have the unfortunate downside [...]
Computational imaging with multiply scattered photons
Abstract: 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 [...]
Mental models for 3D modeling and generation
Abstract: Humans have extraordinary capabilities of comprehending and reasoning about our 3D visual world. One particular reason is that when looking at an object or a scene, not only can we see the visible surface, but we can also hallucinate the invisible parts - the amodal structure, appearance, affordance, etc. We have accumulated thousands of [...]
Complete Codec Telepresence
Abstract: Imagine two people, each of them within their own home, being able to communicate and interact virtually with each other as if they are both present in the same shared physical space. Enabling such an experience, i.e., building a telepresence system that is indistinguishable from reality, is one of the goals of Reality Labs [...]
R.I.P ohyay: experiences building online virtual experiences during the pandemic: what works, what hasn’t, and what we need in the future
Abstract: During the pandemic I helped design ohyay (https://ohyay.co), a creative tool for making and hosting highly customized video-based virtual events. Since Fall 2020 I have personally designed many online events: ranging from classroom activities (lectures, small group work, poster sessions, technical papers PC meetings), to conferences, to virtual offices, to holiday parties involving 100's [...]
Physics-informed image translation
Abstract: Generative Adversarial Networks (GANs) have shown remarkable performances in image translation, being able to map source input images to target domains (e.g. from male to female, day to night, etc.). However, their performances may be limited by insufficient supervision, which may be challenging to obtain. In this talk, I will present our recent works [...]
Weak Multi-modal Supervision for Object Detection and Persuasive Media
Abstract: The diversity of visual content available on the web presents new challenges and opportunities for computer vision models. In this talk, I present our work on learning object detection models from potentially noisy multi-modal data, retrieving complementary content across modalities, transferring reasoning models across dataset boundaries, and recognizing objects in non-photorealistic media. While the [...]
Learning Visual, Audio, and Cross-Modal Correspondences
Abstract: Today's machine perception systems rely heavily on supervision provided by humans, such as labels and natural language. I will talk about our efforts to make systems that, instead, learn from two ubiquitous sources of unlabeled data: visual motion and cross-modal sensory associations. I will begin by discussing our work on creating unified models for [...]
Towards a formal theory of deep optimisation
Abstract: Precise understanding of the training of deep neural networks is largely restricted to architectures such as MLPs and cost functions such as the square cost, which is insufficient to cover many practical settings. In this talk, I will argue for the necessity of a formal theory of deep optimisation. I will describe such a [...]
Towards Interactive Radiance Fields
Abstract: Over the last years, the fields of computer vision and computer graphics have increasingly converged. Using the exact same processes to model appearance during 3D reconstruction and rendering has shown tremendous benefits, especially when combined with machine learning techniques to model otherwise hard-to-capture or -simulate optical effects. In this talk, I will give an [...]
Enabling Self-sufficient Robot Learning
Abstract: Autonomous exploration and data-efficient learning are important ingredients for helping machine learning handle the complexity and variety of real-world interactions. In this talk, I will describe methods that provide these ingredients and serve as building blocks for enabling self-sufficient robot learning. First, I will outline a family of methods that facilitate active global exploration. [...]
How Computer Vision Helps – from Research to Scale
Abstract: Vasudevan (Vasu) Sundarababu, SVP and Head of Digital Engineering, will cover the topic: ‘How Computer Vision Helps – from Research to Scale’. During his time, Vasu will explore how Computer Vision technology can be leveraged in-market today, the key projects he is currently leading that leverage CV, and the end-to-end lifecycle of a CV initiative - [...]
Motion Matters in the Metaverse
Abstract: Abstract: In the early 1970s, Psychologists investigated biological motion perception by attaching point-lights to the joints of the human body, known as ‘point light walkers’. These early experiments showed biological motion perception to be an extreme example of sophisticated pattern analysis in the brain, capable of easily differentiating human motions with reduced motion cues. Further [...]
What do generative models know about geometry and illumination?
Abstract: Generative models can produce compelling pictures of realistic scenes. Objects are in sensible places, surfaces have rich textures, illumination effects appear accurate, and the models are controllable. These models, such as StyleGAN, can also generate semantically meaningful edits of scenes by modifying internal parameters. But do these models manipulate a purely abstract representation of the [...]
Robot Learning by Understanding Egocentric Videos
Abstract: True gains of machine learning in AI sub-fields such as computer vision and natural language processing have come about from the use of large-scale diverse datasets for learning. In this talk, I will discuss if and how we can leverage large-scale diverse data in the form of egocentric videos (first-person videos of humans conducting [...]
From Videos to 4D Worlds and Beyond
Abstract: Abstract: The world underlying images and videos is 3-dimensional and dynamic, i.e. 4D, with people interacting with each other, objects, and the underlying scene. Even in videos of a static scene, there is always the camera moving about in the 4D world. Accurately recovering this information is essential for building systems that can reason [...]
Generative and Animatable Radiance Fields
Abstract: Generating and transforming content requires both creativity and skill. Creativity defines what is being created and why, while skill answers the question of how. While creativity is believed to be abundant, skill can often be a barrier to creativity. In our team, we aim to substantially reduce this barrier. Recent Generative AI methods have simplified the problem for 2D [...]
Generative modeling: from 3D scenes to fields and manifold
Abstract: In this keynote talk, we delve into some of our progress on generative models that are able to capture the distribution of intricate and realistic 3D scenes and fields. We explore a formulation of generative modeling that optimizes latent representations for disentangling radiance fields and camera poses, enabling both unconditional and conditional generation of 3D [...]
Estimating Robustness using Proxies
ABSTRACT: This talk covers some of our recent explorations on estimating the robustness of black-box machine learning models across data subpopulations. In other words, if a trained model is uniformly accurate across different types of inputs, or if there are significant performance disparities affecting the different subpopulations. Measuring such a characteristic is fairly straightforward if [...]
Latent-NeRF for Shape-Guided Generation of 3D Shapes and Textures
Abstract: In this talk, I will focus on presenting my recent work which will be presented at CVPR in less than two months. Text-guided image generation has progressed rapidly in recent years, inspiring major breakthroughs in text-guided shape generation. Recently, it has been shown that using score distillation, one can successfully text-guide a NeRF model to [...]
Navigating to Objects in the Real World
Abstract: Semantic navigation is necessary to deploy mobile robots in uncontrolled environments like our homes, schools, and hospitals. Many learning-based approaches have been proposed in response to the lack of semantic understanding of the classical pipeline for spatial navigation, which builds a geometric map using depth sensors and plans to reach point goals. Broadly, end-to-end [...]
Going Beyond Continual Learning: Towards Organic Lifelong Learning
Abstract: Supervised learning, the harbinger of machine learning over the last decade, has had tremendous impact across application domains in recent years. However, the notion of a static trained machine learning model is becoming increasingly limiting, as these models are deployed in changing and evolving environments. Among a few related settings, continual learning has gained significant [...]
Predictive Scene Representations for Embodied Visual Search
Abstract: My research advances embodied AI by developing large-scale datasets and state-of-the-art algorithms. In my talk, I will specifically focus on the embodied visual search problem, which aims to enable intelligent search for robots and augmented reality (AR) assistants. Embodied visual search manifests as the visual navigation problem in robotics, where a mobile agent must efficiently navigate [...]
Generating Beautiful Pixels
Abstract: In this talk, I will present three experiments that use low-level image statistics to generate high-resolution detailed outputs. In the first experiment, I will use 2D pixels to efficiently mine hard examples for better learning. Simply biasing ray sampling towards hard ray examples enables learning of neural fields with more accurate high-frequency detail in less [...]
Towards Reliable Computer Vision Systems
Abstract: The real world has infinite visual variation – across viewpoints, time, space, and curation. As deep visual models become ubiquitous in high-stakes applications, their ability to generalize across such variation becomes increasingly important. In this talk, I will present opportunities to improve such generalization at different stages of the ML lifecycle: first, I will [...]
Vision without labels
Abstract: Deep learning has revolutionized all aspects of computer vision, but its successes have come from supervised learning at scale: large models trained on ever larger labeled datasets. However this reliance on labels makes these systems fragile when it comes to new scenarios or new tasks where labels are unavailable. This is in stark contrast to [...]
Large Multimodal (Vision-Language) Models for Image Generation and Understanding
Abstract: Large Language Models and Large Vision Models, also known as Foundation Models, have led to unprecedented advances in language understanding, visual understanding, and AI. In particular, many computer vision problems including image classification, object detection, and image generation have benefited from the capabilities of such models trained on internet-scale text and visual data. In [...]
Imaginative Vision Language Models: Towards human-level imaginative AI skills transforming species discovery, content creation, self-driving cars, and emotional health
Abstract: Most existing AI learning methods can be categorized into supervised, semi-supervised, and unsupervised methods. These approaches rely on defining empirical risks or losses on the provided labeled and/or unlabeled data. Beyond extracting learning signals from labeled/unlabeled training data, we will reflect in this talk on a class of methods that can learn beyond the vocabulary [...]
World Knowledge in the Time of Large Models
Abstract: This talk will discuss the massive shift that has come about in the vision and ML community as a result of the large pre-trained language and language and vision models such as Flamingo, GPT-4, and other models. We begin by looking at the work on knowledge-based systems in CV and robotics before the large model [...]