Bharath Hariharan

I am an assistant professor in Computer Science at Cornell University. I work on computer vision and machine learning, in particular on important problems that defy the "Big Data" label. I enjoy problems that require marrying advances in machine learning with insights from computer vision, geometry and domain-specific knowledge. Currently, my group is working on building systems that can learn about tens of thousands of visual concepts with very little or no supervision, produce rich and detailed outputs such as precise 3D shape, and reason about the world and communicate this reasoning to humans. A sampling of the research problems my group works on is presented below; an exhaustive list of publications is available on scholar.


My work has been recognized with an NSF CAREER award and a PAMI Young Researcher Award.

My CV is here and my research statement is here.



Note to prospective PhD students: Admissions at Cornell are done through a committee. If you are interested in working with me, please directly apply through the application website and mention my name

Assistant Professor
311 Gates Hall
Cornell University
bharathh-AT-cs-DOT-cornell-DOT-edu

Teaching

PhD students

Former PhD students

Research

Recognition with minimal labels

Deep learning and ConvNets revolutionized visual recognition, but require large labeled datasets for training. This is a problem in new domains like satellite imagery, in expert applications like fine-grained recognition, and in "open-world" settings like robotics where the space of possible classes is not known a priori. We are designing new classes of recognition systems that can be trained with very few labeled examples and can even discover classes on their own. The key insight is to look beyond the available data, leveraging domain knowledge and visual learning that transcends domains. Funding: This work is funded by an NSF CAREER award, DARPA and IARPA.
Papers

Learning to reconstruct and synthesize 3D

Humans can not just recognize objects but also reconstruct 3D objects. We can reason about 3D even when we can only see one view, or a few sparse views. For computer vision systems to have this ability, they must reason not just about the well-explored geometry of perspective projection, but also about priors about scenes and shapes. Combining mathematical constraints from geometry with the data-driven priors provided by machine learning is an open research question.
Papers

Recognition in 3D

With advances in 3D reconstruction and recognition, vision is now being deployed in a variety of robotics applications, including self-driving cars. These robots have multiple cameras and LiDAR sensors, and require precise 3D location estimates for control. We are bringing insights from recognition, limited-label learning and 3D reconstruction/synthesis to perception in 3D . Funding:This work is funded by NSF.
Papers

Recent papers

New Pre-prints

Recent publications

  • Dynamo-Depth: Fixing Unsupervised Depth Estimation for Dynamical Scenes
    Yihong Sun, Bharath Hariharan
    In NeurIPS, 2023
    TLDR : We improve unsupervised monocular depth estimation for dynamical scenes by modeling 3D independent flow and motion segmentation.
    pdf    bibtex
  • Tracking Everything Everywhere All at Once
    Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, Noah Snavely
    In ICCV, 2023 (Oral)[Best student paper]
    TLDR : A neural field approach to estimating long-term correspondences even through occlusions
    pdf    bibtex
  • Emergent Correspondence from Image Diffusion
    Luming Tang*, Menglin Jia*, Qianqian Wang*, Cheng Phoo, Bharath Hariharan
    In NeurIPS, 2023
    TLDR : Semantic and geometric correspondences can be extracted from diffusion models without any further training
    pdf    bibtex
  • Distilling from Similar Tasks for Transfer Learning on a Budget
    Kenneth Borup, Cheng Perng Phoo, Bharath Hariharan
    In ICCV, 2023
    TLDR : We train models for new domains with limited data and limited compute by identifying pre-training domains and distilling from them
    pdf    bibtex
  • Doppelgangers: Learning to Disambiguate Images of Similar Structures
    Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, Noah Snavely
    In ICCV, 2023 (Oral)
    TLDR : We learn to classify whether an image pair in an SfM pipeline captures the same structure, or similar but distinct structures
    pdf    bibtex
  • Change-Aware Sampling and Contrastive Learning for Satellite Images
    Utkarsh Mall, Bharath Hariharan, Kavita Bala
    In CVPR, 2023
    TLDR : A self-supervised learning approach for satellite images that pushes apart images taken from the same location but over long time spans
    pdf    bibtex
  • Change Event Dataset for Discovery from Spatio-temporal Remote Sensing Imagery
    Utkarsh Mall, Bharath Hariharan, Kavita Bala
    In NeurIPS (Datasets and Benchmarks track), 2022
    TLDR : A new benchmark on discovering meaningful multi-step change events from satellite images with few/no labels
    pdf    bibtex
  • Polynomial Neural Fields for Subband Decomposition and Manipulation
    Guandao Yang*, Sagie Benaim*, Varun Jampani, Kyle Genova, Jonathan T. Barron, Thomas Funkhouser, Bharath Hariharan, Serge Belongie
    In NeurIPS, 2022
    TLDR : A new architecture for neural fields that explicitly decomposes signals into Fourier and other bases
    pdf    bibtex
  • Unsupervised Adaptation from Repeated Traversals for Autonomous Driving
    Yurong You*, Cheng Perng Phoo*, Katie Z Luo*, Travis Zhang, Wei-Lun Chao, Bharath Hariharan, Mark Campbell, Kilian Q. Weinberger
    In NeurIPS, 2022
    TLDR : Data captured from repeated traversals of a scene can be used to adapt 3D object detectors to new domains without supervision
    pdf    bibtex
  • Visual Prompt Tuning
    Menglin Jia*, Luming Tang*, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, Ser-Nam Lim
    In ECCV, 2022
    TLDR : Learning a prompt is the best way to transfer vision transformers to new tasks
    pdf    bibtex
  • Exploiting Playbacks in Unsupervised Domain Adaptation for 3D Object Detection
    Yurong You*, Carlos Andres Diaz-Ruiz*, Yan Wang, Wei-Lun Chao, Bharath Hariharan, Mark Campbell, Kilian Weinberger
    In ICRA, 2022
    TLDR : Extrapolating tracks of detected objects yields good ground truth for adapting 3D detectors
    pdf    bibtex
  • Learning to Detect Mobile Objects from LiDAR Scans Without Labels
    Yurong You*, Katie Luo*, Cheng Perng Phoo, Wei-Lun Chao, Wen Sun, Bharath Hariharan, Mark Campbell, Kilian Weinberger
    In CVPR, 2022
    TLDR : Can we train a 3D object detector without labels by simply driving around?
    pdf    bibtex
  • Hindsight is 20/20: Leveraging past traversals to aid 3D perception
    Yurong You, Katie Luo, Xiangyu Chen, Junan Chen, Wei-Lun Chao, Wen Sun, Bharath Hariharan, Mark Campbell, Kilian Weinberger
    In ICLR, 2022
    TLDR : Can past traversals through a dynamic scene help 3D perception of the current scene?
    pdf    bibtex