Mehmet Kerem Turkcan

Columbia University

New York, NY

I build and deploy AI systems for settings where latency, reliability, and real-world context matter, from city streets and surgical training workflows to complex scientific data. Across these domains, my work turns deep learning research into practical computer vision, multimodal, and data systems that can be tested outside the lab.

At the Center for Smart Streetscapes (CS3) and Civil Engineering & Engineering Mechanics at Columbia University, I lead machine learning efforts that move from research prototypes to fielded systems, including real-time video analytics, open-vocabulary perception, edge and vision-language model workflows, street-scale data pipelines, synthetic data generation, urban digital twins, and foundation model workflows for medical robotics. Before my current position, I was a postdoc at the AIDL Lab in the Department of Electrical Engineering at Columbia University.

DART detections of pedestrians, a bus, and a taxi in Times Square at night

DART Open-vocabulary detection at deployment speed

UrbanOmniDetect 3D cuboids on cars, a bus, and pedestrians seen from above a New York avenue

UrbanOmniDetect View-agnostic urban detection

Surgical SAM 3.1 masks of two graspers and a hook in a laparoscopic frame

Medical Robotics Open datasets and world models for training

Work and leadership

Fielded AI systems: I lead multimodal perception projects for sensing models intended to scale across 900+ New York City intersections and CS3’s three urban testbeds: COSMOS PAWR in New York City, DataCity in New Brunswick, and FAU MobIntel in West Palm Beach.
Work at leading venues: Publications and preprints across CVPR, ICML, ACM UIST, ACM/IEEE SEC, IEEE INFOCOM, IEEE PerCom, EDM, eLife, and medical AI venues.
Open-source systems: Public AI systems including DART and generative-agents, with 1,200+ combined GitHub stars.
Sponsored research: Proposal development, technical reporting, sponsor reviews, and engineering delivery across approximately $30M in institutional research supported by NSF, DARPA, AFOSR, and Con Edison, with additional project support from NVIDIA and EmpireAI.
Teaching and mentorship: Graduate deep learning courses at Columbia, mentoring of Master’s and high school researchers, and service as the main engineering instructor for the CS3 Research Experience for Teachers in 2024, 2025, and 2026.

My current systems include DART for real-time open-vocabulary detection, UrbanOmniDetect and UrbanOmniView for calibration-free monocular 3D perception, PAVE and bikeped for urban safety, city-scale traffic analysis, Loom for analytical neural computing, and world models and open datasets for medical robotics.

Earlier work on language systems includes GPTune, a GPT-2 fine-tuning toolkit from the early public LLM era, and FlyBrainLab, an early prototype of today’s research-oriented AI workbenches. FlyBrainLab’s natural language interface coordinated access to papers, programmable ontologies, and an OrientDB-backed connectome knowledge graph, linking retrieved evidence to large-scale queries, GPU simulation, and interactive 3D visualization. Although it predated general-purpose LLMs, the platform addressed the grounding, retrieval, tool integration, and provenance problems now central to agentic scientific systems.

My independent game and film work also feeds back into my research practice. Through Wisedawn and KEDIKAT, I work with real-time engines, visual storytelling, and production constraints that inform simulation work such as Boundless. KEDIKAT’s film FLEA received four festival accolades in 2026, including Best Storytelling Runner-Up at the MetaMorph AI Award.

News

Oct 6, 2026	Our paper, Harnessing Floating Car Data, Traffic Camera Observations, and Network Flow Analysis for Traffic Volume Estimation, led by Antonina Kosikova, has been accepted for publication in Discover Civil Engineering. The preprint is available on arXiv.
Sep 28, 2026	I will present our vision-based analysis of congestion pricing in New York City, built on 910 public traffic cameras, at the 2026 PAWR PI Meeting. The project was initially supported by compute from the PAWR COSMOS testbed at Columbia University.
Sep 24, 2026	I released BoundlessNYC, a city-scale digital twin of New York City compiled from public municipal records. It covers Manhattan, the Bronx, Brooklyn, and Queens with simulated traffic and pedestrians, pixel-exact segmentation and depth ground truth, and a CARLA-style Python API. An interactive demo runs in the browser.
Sep 23, 2026	I released UrbanAnonymizer, a face and license plate anonymizer for street-level images and video. It outperforms EgoBlur Gen2 on the UFDD and PP4AV benchmarks, and offline tracking keeps masks in place in frames where detections drop out. The models and dataset are on Hugging Face.
Sep 22, 2026	I released StreetPrompt, an open-vocabulary video classifier for street cameras. Users describe situations such as flooding, fog, or collisions in plain text, with no labeled data or retraining, and the models run on hardware from a Jetson Orin Nano to desktop GPUs. The Synthetic Street Scenes training dataset is also public.
Sep 16, 2026	Our Constellation paper, Constellation Dataset: Benchmarking High-Altitude Object Detection for an Urban Intersection, has been published open access in the International Journal of Computer Vision.
Sep 9, 2026	I released UrbanOmniDetect-2, a single network for COCO-80 object detection and calibration-free 3D cuboid estimation from ego-vehicle, infrastructure, and aerial cameras. It ships in five model scales with a real-time tracking and bird’s-eye-view pipeline. Details are on the project page.
Aug 3, 2026	With Soham Samal and Zoran Kostic, I released Surgical SAM 3.1, a text-prompted detector and instance segmenter for surgical instruments, anatomy, and tissue, fine-tuned from SAM 3.1 on four public surgical datasets. The accompanying dataset is also available.

Selected Publications

Calibration-Free View-Agnostic Monocular 3D Object Detection for Urban Scenes

Mehmet Kerem Turkcan, Devika Gumaste, and Zoran Kostic

In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, Jun 2026

HTML PDF Code Dataset Model

UrbanOmniDetect is a calibration-free monocular 3D object detection framework for heterogeneous urban cameras across ego-vehicle, infrastructure, and aerial viewpoints.
AI-Powered CPS-Enabled Vulnerable-User-Aware Urban Transportation Digital Twin: Methods and Applications

Yongjie Fu, Mehmet Kerem Turkcan, Mahshid Ghasemi, Zhaobin Mo, Chengbo Zang, Abhishek Adhikari, Zoran Kostic, Gil Zussman, and Xuan Di

IEEE Transactions on Intelligent Transportation Systems, May 2026

arXiv DOI HTML PDF

This article presents an AI-powered cyber-physical digital twin architecture for vulnerable-user-aware urban transportation, integrating sensing, object detection and tracking, prediction, decision support, and low-latency communications.
Open-H-Embodiment: A Large-Scale Dataset for Enabling Foundation Models in Medical Robotics

Nigel Nelson, Juo-Tung Chen, Jesse Haworth, Xinhao Chen, Lukas Zbinden, Dianye Huang, Mattia Ballo, Filippo Filicori, Mehmet Kerem Turkcan, and others

Apr 2026

arXiv DOI HTML Code Dataset Model

Open-H-Embodiment is a multi-institution open dataset for medical robot learning with synchronized video and kinematics across healthcare robotic platforms, enabling foundation models and world models for medical robotics.
Detect Anything in Real Time: From Single-Prompt Segmentation to Multi-Class Detection

Mehmet Kerem Turkcan

Mar 2026

arXiv DOI Code Model

DART is a training-free framework that converts a promptable frontier vision model into a real-time multi-class open-vocabulary detector by sharing class-agnostic visual backbone computation, batching decoding, and optimizing inference for deployment.
Real-Time Video Analytics for Urban Safety: Deployment over Edge and End Devices

Mahshid Ghasemi, Yongjie Fu, Xinyu Ouyang, Peiran Wang, Mehmet Kerem Turkcan, Jhonatan Tavori, Sofia Kleisarchaki, Thomas Calmant, Levent Gurgen, Zoran Kostic, Xuan Di, Gil Zussman, and Javad Ghaderi

In Proceedings of the Tenth ACM/IEEE Symposium on Edge Computing, 2025

Best Paper Award

DOI HTML

PAVE is a privacy-preserving real-time video analytics system that uses edge and end-device processing to support pedestrian safety at urban intersections.
Accelerating with FlyBrainLab the discovery of the functional logic of the Drosophila brain in the connectomic and synaptomic era

Mehmet Kerem Turkcan, Aurel A. Lazar, Tingkai Liu, and Yiyin Zhou

Elife, Mar 2021

DOI HTML Code Dataset Website

In recent years, a wealth of Drosophila neuroscience data have become available including cell type and connectome/synaptome datasets for both the larva and adult fly. To facilitate integration across data modalities and to accelerate the understanding of the functional logic of the fruit fly brain, we have developed FlyBrainLab, a unique open-source computing platform that integrates 3D exploration and visualization of diverse datasets with interactive exploration of the functional logic of modeled executable brain circuits. FlyBrainLab’s User Interface, Utilities Libraries and Circuit Libraries bring together neuroanatomical, neurogenetic and electrophysiological datasets with computational models of different researchers for validation and comparison within the same platform. Seeking to transcend the limitations of the connectome/synaptome, FlyBrainLab also provides libraries for molecular transduction arising in sensory coding in vision/olfaction. Together with sensory neuron activity data, these libraries serve as entry points for the exploration, analysis, comparison, and evaluation of circuit functions of the fruit fly brain.