A published research team, and the products that come out of it
Cyclope.ai's research is led by Amir Nakib and relies on a network of partner laboratories - CRIStAL (CNRS), Inria Lille, CentraleSupélec, LISSI. What is published in leading journals and conferences is what runs in production on the network.
- ICCV, IEEE T-ITSleading venues in computer vision
- 4 laboratoriesCRIStAL·CNRS, Inria Lille, CentraleSupélec, LISSI
- 2024road safety innovation award (Patrolcare)
Research areas
Each area addresses a production constraint: transaction-level accuracy, degraded environments, data protection.
Fine-grained classification
Vehicle class recognition where a 1% error is a direct economic loss.
Scene understanding
Semantic segmentation of infrastructure: lanes, hard shoulders, sidewalks, tunnels.
Privacy and robustness
Image anonymization without loss of accuracy, robustness to perturbations.
Optimization and frugality
Learning from less annotated data, running inference on constrained hardware.
Scientific publications
A selection of work published by the team since 2019, with Cyclope.ai affiliation.
Every year
at least one publication since 2019, without interruption
1,800+
citations of the team's work (Google Scholar)
2023
PrivacyAdaptive Image Anonymization in the Context of Image Classification with Neural Networks
N. Shvai, A. Llanza Carmona, A. Nakib - ICCV
Anonymizing images (faces, license plates) without degrading the accuracy of classification models in production.
2020
ClassificationAccurate Classification for Automatic Vehicle-Type Recognition Based on Ensemble Classifiers
N. Shvai, A. Hasnat, A. Meicler, A. Nakib - IEEE Trans. on Intelligent Transportation Systems
The vehicle classification method at the heart of Tollsense: fusion of CNNs and optical sensors through gradient boosting.
2019
ClassificationNovel Context-Aware Classification for Highly Accurate Automatic Toll Collection
M. Khata, N. Shvai, A. Llanza, A. Sanogo, A. Meicler, A. Nakib - IEEE Intelligent Vehicles Symposium (IV)
Using the spatial relationships between objects in the scene to classify with 74 times less annotated data.
2019
SegmentationSemantic Segmentation Approach for Tunnel Roads’ Analysis
A. Llanza, A. Sanogo, M. Khata, K. Alami, N. Shvai, A. Hasnat, A. Meicler, Y. El Khattabi, J. Noslier, P. Maarek, A. Nakib - SPIE Applications of Digital Image Processing XLII
Segmenting lanes, hard shoulders and sidewalks in tunnels - the basis of Tunnelwatch's incident localization.
Full list on Prof. Amir Nakib's Google Scholar profile.
Filed patents
Two protected technologies, both in production.
- PatentToll classification
Vehicle classification through vision and sensor fusion
The patented method at the heart of Tollsense: the class of each vehicle determined without contact, verified on 1,200 lanes in production.
- PatentCarpool lane enforcement
Vision-based vehicle occupancy counting
The patented technology behind Roadshare: the number of people on board determined on the fly, through the windshield, day and night.
From publication to product
Ensemble classification
IEEE T-ITS 2020 → classification engine of Tollsense, 1,200 lanes in production.
Tunnel segmentation
SPIE 2019 → zone-based incident localization in Tunnelwatch, ~540 cameras.
Adaptive anonymization
ICCV 2023 → privacy-preserving processing in Travelmatch and across the platform.
Work with the research team
Industrial PhDs, academic collaborations, evaluations on your data: get in touch.