Recent Results

  • [2024] Our work on shape-based measures for scene categorization is now available as a full length article in PAMI here.
  • [2023] Our work on recovering cardiomyocyte orientation at the micron scale made the cover page of the October issue in EMBO, which is here. See also the Faculty of Science piece here and the promotional video.
  • [2023] Our work on ultrastructure analysis of astrocytes is now out in Current Biology [PDF]. See Chris Salmon's twitter thread , the RI-MUHC article and Joselyn Soto and Baljit Khakh's Current Biology dispatch article.
  • [2022] Arnab's work on equivariant representation learning, joint with Siamak's group, will be presented at ICML 2022. [PDF ]
  • [2022] Morteza's work on medial spectral coordinates for computer vision problems was presented as a talk at CVPR 2022. [arXiv]
  • [2021] Working closely with Minhaj Sirajuddin's group at INSTEM-NCBS, we have new discoveries which relate to the geometry of heart wall fibers in the mammalian heart. Stay tuned!
  • [2021] Peter's work using mean curvature for feature computation appeared in Radiology: AI with an insightful commentary by Michael Vannier, pointing to the need to return to classical scale-space methods in radiology.
  • [2020] Chu and Babak's work on affinity graph supervision appeared in CVPR 2020. Here we showed that learning affinity weights between entities can benefit a variety of vision tasks. An arXiv version of the paper is here.
  • [2020] Charles-Olivier and Morteza's work also appeared in CVPR 2020. Here we showed that an algorithm that explicitly uses the rules of the shock grammar (introduced in my doctoral work) provides an effective method for medial axis extraction from images without leaqrning. An arXiv version of the paper is here.
  • [2019] Morteza and Gabriel's work appeared in CVPR 2019. Here we showed that medial axis based salience measures which capture local mirror symmetry aid scene categorization from contours. The paper is available here.
  • [2019] Work in collaboration with Wang, Xu, Bai, Tsogkas and Dickinson appeared at CVPR 2019. Here we showed that our flux based representations could be used by a deep network to achieve state of the art performance in medial axis detection from images, with learning. The paper is available here and an extended version is now published in IJCV.
  • [2018] In our Scientific Reports, paper we develop the argument that the particular helicoidal myofiber geometry we reported in our PNAS paper in fact minimizes diffusion bias. Hence, this geometry apparently facilitates both heart wall mechanics and conduction.

    • Some Past Contributions

  • The shock graph, which has lead to many papers on graph-based object recognition.
  • The Hamilton-Jacobi skeleton, a robust flow-based approach to computing medial representations.
  • Flux invariants for describing 2D and 3D forms.
  • Flux maximizing flows for segmentation.
  • Our "Medial Representations" book (Siddiqi and Pizer, Springer 2008), which serves as a graduate level text on this subject.
  • Our discovery that heart wall myofibers are organized locally in minimal surfaces that are generalized helicoids.
  • New discoveries, with Keith Murai's group, related to astrocyte ultrastructure at the nanometer scale.