Or Hirschorn

Or Hirschorn

I am a Ph.D. candidate at the School of Electrical Engineering at Tel-Aviv University, under the joint supervision of Prof. Shai Avidan.

Currently interning at Amazon Prime Video working to extend 3D reconstruction quality using generative models.

My research interests include machine learning and computer vision. More specifically, More specifically, I am interested in Foundation Models, Semantic Correspondence, and 3D representations, with a particular focus on diffusion-based methods for 3D reconstruction.

Email Github Google Scholar Twitter LinkedIn

Papers Talks

Publications

SpheRoPE: Zero-Shot Optimization-Free 360° Panorama Generation with Spherical RoPE

UNDER REVIEW

Or Hirschorn, Aaron Olender, Eli Alshan, Ianir Ideses, Lior Fritz, Sagie Benaim

We present a zero-shot, training-free and optimization-free framework for generating 360° panoramic images and videos by directly injecting spherical priors into pre-trained diffusion transformers. Our Spherical RoPE re-parameterizes low-frequency positional channels as 3D Cartesian coordinates and harmonically quantizes high-frequency channels, enforcing exact periodicity and polar convergence while preserving pretrained local priors.

Project Page arXiv Code

Multi-View Foundation Models teaser

Multi-View Foundation Models

ECCV 2026

Or Hirschorn*, Leo Segre*, Shai Avidan

We introduce multi-view foundation models that equip powerful 2D vision backbones with the ability to reason jointly across multiple camera views. By explicitly incorporating camera geometry, our approach produces 3D-consistent semantic representations without per-scene optimization, enabling scalable 3D understanding and correspondence.

arXiv Code

Splatent: Splatting Diffusion Latents for Novel View Synthesis

CVPR 2026

Or Hirschorn, Omer Sela, Inbar Huberman-Spiegelglas, Netalee Efrat, Eli Alshan, Ianir Ideses, Frederic Devernay, Frederic Devernay, Yochai Zvik, Lior Fritz

We show that the VAE space of current diffusion models is not 3D consistent. To recover details lost in 3D reconstruction, we suggest a 2D diffusion based method to restore lost information.

arXiv Code

Edge Weight Prediction For Category-Agnostic Pose Estimation

ICLR 2026

Or Hirschorn, Shai Avidan

We perform category-agnostic pose estimation by predicting structural edge weights to improve keypoint localization accuracy.

arXiv Code

CapeX: Category-Agnostic Pose Estimation from Textual Point Explanation

ICLR 2025

Matan Rusanovsky, Or Hirschorn, Shai Avidan

We perform open-vocabulary pose estimation. Given text prompt describing object parts, our method localizes these keypoints in an input image.

arXiv Code

Sifting through the Haystack - Efficiently Finding Rare Animal Behaviors in Large-Scale Datasets

WACV 2025

Shir Bar, Or Hirschorn, Roi Holzman, Shai Avidan

We present a pipeline leveraging anomaly detection with Graph Convolutional Networks to efficiently sample rare animal behaviors from large-scale, unlabeled datasets, significantly reducing annotation costs while improving classifier performance, particularly for behaviors as rare as 0.02% of the data.

arXiv Code

A Graph-Based Approach for Category-Agnostic Pose Estimation

ECCV 2024

Or Hirschorn, Shai Avidan

Pose Anything - given only one example image and skeleton, our method can perform pose estimation on unseen categories.

Project Page arXiv Code

Optimize and Reduce: A Top-Down Approach for Image Vectorization

AAAI 2024

Or Hirschorn*, Amir Jevnisek*, Shai Avidan

We propose Optimize & Reduce (O&R), a top-down approach to vectorization that is both fast and domain-agnostic. O&R aims to attain a compact representation of input images by iteratively optimizing Bézier curve parameters and significantly reducing the number of shapes, using a devised importance measure.

Project Page arXiv Code

Normalizing Flows for Human Pose Anomaly Detection

ICCV 2023

Or Hirschorn, Shai Avidan

We perform video anomaly detection, distilling the problem to anomaly detection of human pose. We leverage the highly compact pose representation in a normalizing flows framework, which we extend to tackle the unique characteristics of spatio-temporal pose data and show its advantages in this use case.

arXiv Code

Shape-Consistent Generative Adversarial Networks for Multi-Modal Medical Segmentation Maps

ISBI 2022

Leo Segre*, Or Hirschorn*, Dvir Ginzburg, Dan Raviv

We use cross-modality adaptation between CT and MRI whole cardiac scans for semantic segmentation.

arXiv Code

Talks

Vision Day 2024

Splatting Diffusion Latents for Novel View Synthesis

Amazon Prime Video Sport Reading Group

December 2, 2025

Graph-Based Category-Agnostic Pose Estimation

The Hebrew University Computer Vision Seminar

March 30, 2025

Graph-Based Category-Agnostic Pose Estimation

Google Research Israel Reading Group

January 7, 2025

Graph-Based Category-Agnostic Pose Estimation

Computer Vision Day Israel

December 16, 2024

Graph-Based Category-Agnostic Pose Estimation

Weizmann Institute of Science Computer Vision Seminar

December 5, 2024

Semantic Vector Representations in the Service of Computer Vision

Google Research Boston Seminar

February 28, 2024

Semantic Vector Representations in the Service of Computer Vision

Adobe Research Seattle Seminar

February 20, 2024

Semantic Vector Representations in the Service of Computer Vision

The Hebrew University Computer Vision Seminar

January 28, 2024

Pose Anything - A Graph-Based Approach to Category-Agnostic Pose Estimation

Tel-Aviv University Computer Vision Seminar

January 16, 2024

Normalizing Flows for Human Pose Anomaly Detection

Tel-Aviv University Computer Vision Seminar

June 14, 2023