Sebastian Houben

[Sebastian Houben @SebHouben]

2023: A student presenting a self-trained 60 million parameter model stating that it is "rather small".

[Sebastian Houben @SebHouben]

"In order to ensure the expeditious review process at the [Journal] I would greatly appreciate it if you could complete your review by June 30, 2008."

Last time I received such a request was from the Journal of Applied Tachyon Containment, ..., or is it "will have received"?!

[Sebastian Houben @SebHouben]

We study the effect that different numbers and configurations of keypoints as well as different neural detection architectures and heuristics (like visibility analysis) have on the door detection performance.

[Sebastian Houben @SebHouben]

@TBarowski, Brehme, Szczot, Houben, Dynamic Door Modeling for Monocular 3D Vehicle Detection, Intelligent Vehicles Symposium 2021

https://t.co/hMwDtx2rsP

We extend keypoint-based pose estimation of vehicles to dynamic parts to detect which doors are open. 🧵

[Sebastian Houben @SebHouben]

With first experimental results demonstrated, we propose an automated feedback loop that identifies and generates data from underrepresented (or missing) situations and adds it to the training regimen.

[Sebastian Houben @SebHouben]

Our paper

Gannamaneni et al., Semantic Concept Testing in Autonomous Driving by Extraction of Object-Level Annotations from CARLA

was accepted at International Conference on Computer Vision, Workshop on Embedded and Real-World Computer Vision in Autonomous Driving

#ICCV2021 🧵

[Sebastian Houben @SebHouben]

We extend CARLA to generate object-level meta-annotations. This allows for systematic testing and analysis of a state-of-the-art pedestrian detection model where change in performance can be attributed to skin color, size, relative position, and clothing.

[Sebastian Houben @SebHouben]

This study is a fruitful collaboration within the competency network Machine Learning Rhine-Ruhr (@MLCompetence) including scientists from the University of Bonn, the Technical University Dortmund and the @FraunhoferIAIS. 🧵2/2

[Sebastian Houben @SebHouben]

In our new survey
@kbeckh et al., "Explainable Machine Learning with Prior Knowledge: An Overview", 2021
https://t.co/p6wQIBAqV9
we elucidate how prior knowledge benefits or begets explainable machine learning and cover recent research that describe ways to achieve this. 🧵1/2

[Sebastian Houben @SebHouben]

The validity of our approach is (true to our trademark) demonstrated on the task of traffic sign classification.

#machinelearning #autonomousdriving #datageneration #iv2021

[Sebastian Houben @SebHouben]

Badly rendered data point or useful piece of data variance?

Horn, Janssen, Houben, "Automated Selection of High-Quality Synthetic Images for Data-Driven Machine Learning: A Study on Traffic Signs", Intelligent Vehicles Symposium 2021

https://t.co/FXhQg3hae4

[Sebastian Houben @SebHouben]

In our paper, we use state-of-the-art uncertainty measures to derive the "better" portion of a synthetic dataset and show that we can do with only 40% to 5% of the data without losing accuracy.

[Fraunhofer IAIS @FraunhoferIAIS]

#OUTNOW: Der #Fraunhofer #KI-Prüfkatalog für vertrauenswürdige #KünstlicheIntelligenz! Mit dem +160 Seiten starken Leitfaden können Unternehmen u. Entwickler*innen KI-Systeme sicher u. verlässlich prüfen u. gestalten.
📖Zum Download: https://t.co/A8kzLysam4
@_KINRW @WirtschaftNRW

[Fraunhofer IAIS @FraunhoferIAIS]

Nicht verpassen! 👉 #Keynote von Dr. Maximilian Poretschkin zu »#TrustworthyAI – Obligation or Entrepreneurial Opportunity?« auf der EAA e-Conference.

🗓️ 29. Juni, 9.40 bis 10.40 Uhr, online

🔗 Infos & Tickets: https://t.co/WM20QQAGqH

#KünstlicheIntelligenz #ZuverlässigeKI

[Sebastian Houben @SebHouben]

We demonstrate how to replace vehicle license plates on natural images maintaining partial control over the result (e.g. the code and perspective). We validate it by systematic testing of a state-of-the-art license plate reading system.

#ITSC2021 #autonomusdriving #AI

[Sebastian Houben @SebHouben]

Our paper

Kacmaz, Melchior, Horn (@Daniela_R_Horn), Witte, Schoenen, Houben, "Fully Automated, Realistic License Plate Substitution in Real-Life Images"

was accepted for publication at the Intelligent Transportation Systems Conference (ITSC).

(no real license plate shown)

[Sebastian Houben @SebHouben]

We propose a framework deploying label-to-image synthesis to inspect to what extent we can transfer testing results of semantic segmentation from synthetic to real-life data.
Have a look!

#IEEE #IV2021 #IV21 #autonomousdriving

[Sebastian Houben @SebHouben]

Our paper was accepted to the Workshop on Ensuring and Validating Safety for Automated Vehicles at the Intelligent Vehicles Symposium

Rosenzweig, Brito et al., Validation of Simulation-Based Testing: Bypassing Domain Shift with Label-to-Image Synthesis

https://t.co/bMIvPzXLCJ

[Sebastian Houben @SebHouben]

Soft logic rules can describe common spatial relationships among objects. We use this to robustify semantic image segmentation.

Our CVPR-SAIAD paper: "Plants Don't Walk on the Street: Common-Sense Reasoning for Reliable Semantic Segmentation"
https://t.co/oofv2g70Dd

#CVPR2021