Title: Cross-modal generation and understanding of multimodal content

In the first part of the presentation, we will present our work on video generation without annotations or prior object-specific information. Trained on videos of similar objects (e.g. faces, bodies), our method generalizes across the category. Building on this, we introduce a Learnable Game Engine (LGE), trained from monocular annotated videos, that maintains scene and object states and renders environments from controllable viewpoints. Like a game engine, it simulates physics and logic, allowing users to control the game play or use a director mode to guide agents via high-level language and goals, enabled by learned game AI. The second part will investigate the safety and fairness of the current generative models. While most of the existing research focuses on detecting closed sets of biases defined a priori, we tackle the challenge of open-set bias detection in text-to-image generative models. For this we proposed OpenBias, a new pipeline that agnostically identifies and quantifies the severity of biases without access to any precompiled set. We study the behavior of Stable Diffusion 1.5, 2, and XL emphasizing new biases, never investigated before. Via quantitative experiments, we demonstrate that OpenBias agrees with current closed-set bias detection methods and human judgement.

Title: Reconstructing, Editing, and Generating the Interactive 4D World

The world around us is 4 dimensional (3D space + time) -- objects, people, machines move in 3D space and across time. Yet, most computer vision and AI methods focus on 2D images/videos, or static 3D scenes. In this talk, I will share some of my group's ongoing work on 4D spatio-temporal understanding. Specifically, I will introduce PackUV, a new representation and method for reconstructing 4D volumetric videos; and MotionSplicer, a method for part-based editing of 4D volumetric videos. I will also introduce GenHSI, a method for 4D human motion generation in scenes; and GigaHands, a large-scale 3D human hand activity dataset that provides visual and contact information for robot manipulation learning. We show that 4D spatio-temporal understanding of the real world is still largely unsolved, and requires addressing significant challenges.

Title: Anticipating the Road Ahead in a Landscape of Disrupted Paths

Camera-based perception has advanced dramatically, with modern deep neural models achieving impressive results across diverse and challenging road environments worldwide. Yet a fundamental question remains open: how do we anticipate what humans (drivers, riders, or pedestrians) intend to do, especially on the unstructured, heterogeneous roads that characterize countries like India? Understanding the unfolding dynamics of human actions from video requires models that go beyond classical computer vision and scene understanding.

Computer vision itself is in a moment of transition. Decades of geometric and photometric reasoning have given way to data-driven learning at scale. This shift prompts deeper reflection: what foundational ideas from classical vision remain essential?How do we cast and solve problems? How do we balance between the novelty of the algorithms, incremental improvements over the state of the art and immediate impact? This talk distills experiences from working on perception and intention-prediction models in Indian traffic—settings that expose the limitations of existing approaches and open new research questions.

Title: Translating AI Innovations into Solutions in the Global South

According to Prof. Kentaro Toyama, technology predominantly acts as an amplifier of human effort and the strength of institutions. In this presentation, I will share examples from India and other developing nations that demonstrate how AI-powered solutions are being designed, implemented, and scaled to improve people's lives, in manner consistent with Kentaro’s observation. Each project narrative will explore how specific problems were selected, the significant constraints and opportunities encountered during the design and development phases, and the approaches taken for implementation. The discussion will also address the challenges and possibilities associated with scaling these solutions. A central lesson from these experiences is the necessity for what I call “radical collaboration.” This involves bringing together a diverse group of contributors—technology specialists, implementing organizations, subject matter experts, donors, and government agencies. Through these case studies, I will illustrate how such deep collaboration can turn AI-driven innovations into effective solutions for challenges in the developing world.