About me

Hi, I am Oussema Dhaouadi, a PhD student at the Technical University of Munich in the Computer Vision Group led by Prof. Daniel Cremers. I am currently an AI Applied Scientist - PhD Intern at Zillow Group. I previously was a visiting researcher at the University of Cambridge in the CV4DT Group led by Dr. Olaf Wysocki, and visited ETH Zürich in the Computer Vision and Geometry Group led by Prof. Marc Pollefeys. I also supervised a research project at TUM.ai in collaboration with the University of Cambridge. During my PhD, I worked in collaboration with DeepScenario on 3D reconstruction and localization.

Before starting my PhD, I was a Machine Learning Engineer at Knowlix GmbH, where I built automated solutions for paperwork. I received my Master’s degree with high distinction in Electrical Engineering and Information Technology from TUM. I completed my thesis at the BMW Group Research and Technology Center under the supervision of Dr. Ahmet Firintepe and Prof. Eckehard Steinbach and spent a semester abroad at NTU Singapore. I also hold a Bachelor’s degree in Electrical Engineering and Information Technology from TUM, where I completed my Bachelor’s thesis under the supervision of Alexander Kuhn and Prof. Alois Knoll. During my studies, I gained additional experience at Fujitsu and Rohde & Schwarz.

News

Publications

PlanarCityLoc: Breaking Silhouette Ambiguity in LoD Aerial Localization via 3D Planes Oussema Dhaouadi, Shao Jie Hu-Chen, Jakob Gollreiter, Samir Jusufi, Nikolas Rieger, Caroline Wendlandt, Christian Wender, Jingfeng Mao, Guangming Wang, Zuria Bauer, Marc Pollefeys, Daniel Cremers, Olaf Wysocki WACV 2027 Project page coming soon. We present PlanarCityLoc, a framework that resolves silhouette ambiguity in aerial localization against textureless LoD city models by matching facade and roof planes between query images and the 3D map, enabling analytical PnP refinement and significantly improving accuracy on UAVD4L-LoDv2 and Swiss-EPFL-v2.

MatchCityLoc: Aerial Localization in LoD City Models via Surface Normal Matching Oussema Dhaouadi, Zuria Bauer, Johannes Meier, Marc Pollefeys, Daniel Cremers, Olaf Wysocki WACV 2027 Project page coming soon. We present MatchCityLoc, a zero-shot UAV localization framework that matches foundation-model surface normals from the query image against LoD-rendered normals to establish explicit 3D-to-2D correspondences and iteratively solve the pose with PnP, achieving state-of-the-art accuracy with very few renderings.

MAGneT-3D: Monocular and Domain-Generalizable Temporal 3D Detection Mohamed Kotb*, Johannes Meier*, Christoph Reich, Oussema Dhaouadi, Luis Denninger, Daniel Cremers ECCV 2026 DriveX Workshop (Oral) Project Page | Paper We introduce MAGneT-3D, the first domain-generalized monocular temporal 3D object detector, combining domain-robust dynamic anchor generation with temporal refinement and identity merging to improve average cross-dataset NDS from 12.1% to 18.6% while also increasing in-domain accuracy.

Combining Projected Uncertainty for Self-Supervised Visual Odometry: From Two-Frame to Multi-Frame Jingchao Xie*, Oussema Dhaouadi*, Weirong Chen, Johannes Meier, Zuria Bauer, Marc Pollefeys, Daniel Cremers IJCV 2026 Project Page | Paper | Github Repo We present CoProU, a principled probabilistic formulation that propagates and fuses uncertainties across temporal frames to effectively identify dynamic regions and temporal inconsistencies, demonstrating its versatility through a two-frame framework (CoProU-VO-2F) and a multi-frame extension (CoProU-VO-MF) with a unified transformer architecture for self-supervised visual odometry.

SemCityLoc: Aerial 6DoF Localization Using Semantic 3D City Models Jingfeng Mao, Xuyang Chen, Qilin Zhang, Oussema Dhaouadi, Guangming Wang, Brian Sheil, Daniel Cremers, Yan Xia, Olaf Wysocki ECCV 2026 Project Page | Paper We propose SemCityLoc, a semantic-geometric alignment system that reframes aerial pose estimation as structured surface registration between foundation-model-derived visual priors and standardized LoD-compliant 3D city models, reducing mean positional error from 9.89m to 2.62m in challenging urban canyons.

OrthoTrack: Continuous 6-DoF UAV Trajectory Estimation via Orthophoto-Anchored Tracking Oussema Dhaouadi, Zuria Bauer, Johannes Meier, Olaf Wysocki, Marc Pollefeys, Daniel Cremers ECCV 2026 Project Page | Paper | Data | Github Repo We present OrthoTrack, a training-free system that estimates continuous 6-DoF UAV trajectories using only publicly available orthophotos and surface models as a map prior, and introduce the MovingDrone Dataset, a large-scale benchmark pairing photorealistic UAV sequences with dense 6-DoF ground truth and co-registered multi-modal geodata.

LeAD-M3D: Leveraging Asymmetric Distillation for Real-Time Monocular 3D Detection Johannes Meier, Jonathan Michel, Oussema Dhaouadi, Yung-Hsu Yang, Christoph Reich, Zuria Bauer, Stefan Roth, Marc Pollefeys, Jacques Kaiser, Daniel Cremers ECCV 2026 Project Page | Paper | Github Repo We present LeAD-M3D, a real-time monocular 3D object detector that combines asymmetric distillation, 3D-aware matching, and confidence-gated inference to improve accuracy and efficiency without LiDAR, stereo, or geometric priors.

IDEAL-M3D: Instance Diversity-Enriched Active Learning for Monocular 3D Detection Johannes Meier, Florian Günther, Riccardo Marin, Oussema Dhaouadi, Jacques Kaiser, Daniel Cremers WACV 2026 Paper We introduce IDEAL-M3D, an instance-level active learning approach for monocular 3D detection that uses diverse model ensembles to prioritize informative objects for annotation. It matches or exceeds full-dataset performance on KITTI using only 60% of the annotations.

GrounDiff: Diffusion-Based Ground Surface Generation from Digital Surface Models Oussema Dhaouadi, Johannes Meier, Jacques Kaiser, Daniel Cremers WACV 2026 Project Page | Paper We introduce GrounDiff, a diffusion-based method that generates DTMs from DSMs by iteratively removing non-ground structures, achieving up to 93% lower error than prior learning-based approaches, and present PrioStitch and GrounDiff+ for scalable high-resolution terrain generation and smooth road reconstruction with up to 81% lower distance error.

OrthoLoC: UAV 6-DoF Localization and Calibration Using Orthographic Geodata Oussema Dhaouadi, Riccardo Marin, Johannes Meier, Jacques Kaiser, Daniel Cremers NeurIPS 2025 (Oral) Project Page | Paper | Data | Github Repo We introduce OrthoLoC, the first large-scale UAV localization dataset with 16,425 images from Germany and the U.S., paired with orthophotos and digital surface models (DSMs) to enable benchmarking under domain shifts, and propose AdHoP, a refinement method that improves feature matching accuracy by up to 95% and reduces translation error by 63%, advancing lightweight high-precision aerial localization without GNSS or large 3D models.

CoProU-VO: Combining Projected Uncertainty for End-to-End Unsupervised Monocular Visual Odometry Jingchao Xie*, Oussema Dhaouadi*, Weirong Chen, Johannes Meier, Jacques Kaiser, Daniel Cremers GCPR 2025 (Oral - Best Paper Award) Project Page | Paper | Github Repo | Poster We propose CoProU-VO, an unsupervised visual odometry framework that propagates and combines uncertainty across temporal frames to robustly filter dynamic objects, leveraging vision transformers to jointly learn depth, uncertainty, and poses, and achieving superior performance on KITTI and nuScenes, especially in challenging dynamic and highway scenarios.

Shape Your Ground: Refining Road Surfaces Beyond Planar Representations Oussema Dhaouadi, Johannes Meier, Jacques Kaiser, Daniel Cremers IV 2025 Project Page | Paper | Data | Poster We present FlexRoad, the first framework for smooth and accurate road surface reconstruction using NURBS fitting with Elevation-Constrained Spatial Road Clustering, and introduce the GeoRoad Dataset (GeRoD) for benchmarking, demonstrating that FlexRoad outperforms existing methods across diverse datasets and terrains while robustly handling noise and input variability.

Highly Accurate and Diverse Traffic Data: The DeepScenario Open 3D Dataset Oussema Dhaouadi*, Johannes Meier*, Luca Wahl, Jacques Kaiser, Luca Scalerandi, Nick Wandelburg, Zhuolun Zhou, Nijanthan Berinpanathan, Holger Banzhaf, Daniel Cremers IV 2025 Project Page | Paper | Data | Poster We introduce DSC3D, a large-scale, occlusion-free 3D trajectory dataset captured via a monocular drone tracking pipeline, featuring over 175,000 trajectories of 14 traffic participant types across diverse urban and suburban scenarios, designed to enhance autonomous driving systems with detailed environmental representations and improved obstacle interaction modeling.

MonoCT: Overcoming Monocular 3D Detection Domain Shift with Consistent Teacher Models Johannes Meier*, Louis Inchingolo*, Oussema Dhaouadi, Yan Xia*, Jacques Kaiser, Daniel Cremers ICRA 2025 (Oral) Paper We address monocular 3D object detection across diverse sensors and environments with MonoCT, an unsupervised domain adaptation method that uses Generalized Depth Enhancement, Pseudo Label Scoring, and Diversity Maximization to generate high-quality pseudo labels, achieving up to 21% improvement over prior state-of-the-art methods across six benchmarks.

TrafficLoc: Localizing Traffic Surveillance Cameras in 3D Scenes Yan Xia*, Yunxiang Lu*, Rui Song, Oussema Dhaouadi, João F. Henriques, Daniel Cremers ICCV 2025 Project Page | Paper We tackle traffic camera localization in 3D maps by introducing TrafficLoc, a coarse-to-fine image-to-point cloud registration method with Geometry-guided Attention Loss, Inter-intra Contrastive Learning, and Dense Training Alignment, alongside the Carla Intersection dataset, achieving up to 86% improvement over prior methods and new state-of-the-art performance on KITTI and NuScenes for both in-vehicle and traffic cameras.

CARLA Drone: Monocular 3D Object Detection from a Different Perspective Johannes Meier, Luca Scalerandi, Oussema Dhaouadi, Jacques Kaiser, Nikita Araslanov, Daniel Cremers GCPR 2024 (Oral) Project Page | Github Repo | Paper | Data We introduce the CARLA Drone (CDrone) dataset to expand 3D detection benchmarks with diverse camera perspectives and propose GroundMix, a ground-based 3D-consistent augmentation pipeline, demonstrating that prior methods struggle across these views while our approach significantly improves detection accuracy across both synthetic and real-world datasets.

Comparing Head and AR Glasses Pose Estimation Ahmet Firintepe, Oussema Dhaouadi, Alain Pagani, Didier Stricker ISMAR-Adjunct 2021 Paper | IEEE We introduce CapsPose, a novel capsule-network-based 6-DoF pose estimator, and compare head pose versus AR glasses pose estimation, showing that AR glasses pose is generally more accurate, while CapsPose significantly outperforms existing methods on the HMDPose dataset.

Other Projects

SeqCapsGAN: Generating Image Stylized Captioning using Capsule GANs Course: Natural Language Processing 2020 Github Repo | Report We introduce SeqCapsGAN, a stylized image captioning framework that combines Generative Adversarial Networks and Capsule Networks to generate human-like captions with sentiment, outperforming baseline models on key evaluation metrics.

EMGCaps: Electromyographic Hand Gesture Recognition using Capsule Network Practical course at NTU 2020 Github Repo | Report We introduce EMG-Caps, a framework for hand gesture classification using Capsule Networks on sEMG signals from the NinaPro DB5 dataset, demonstrating that CapsNets outperform conventional classifiers and other deep learning methods—achieving up to 93.27% accuracy across 53 hand movements.

Electromyography-based Control of a Simulated Robot Arm Practical course Networked and Cooperative Systems by Prof. Sandra Hirche 2019 Github Repo | Report We performed data acquisition and preprocessing, designed and evaluated classification systems using both conventional and neural network–based pattern recognition methods in TensorFlow and Keras, and implemented a real-time operating 6-DoF robot arm.

Trajectory Planning for Humanoid RoboCup using bi-RTT Algorithm Practical course Advanced Lab Humanoid RoboCup by Prof. Gordon Cheng 2019 Github Repo We implemented trajectory planning algorithms (RRT/RRT*) in C++, simulated and evaluated their performance in both simple and challenging scenarios, and conducted real-world testing on the NAO robot using ROS.

Hand Gesture Controlled Obstacle Avoiding Robot Project of the course Humanoid Sensors and Actuators by Prof. Gordon Cheng 2019 Github Repo We developed Hand Gesture Controlled Obstacle Avoiding Robot, implementing embedded programming in C++ on an Atmega328P microcontroller, sensor signal processing with a 6DoF IMU, ultrasonic and IR sensors, actuation control for servos, gear motors, and vibration motors, and a custom communication protocol for a radio transmitter/receiver module.

Prediction Opening-Weekend Box Office Sales based on Movie Trailers Project of the course Applied Machine Learning by Prof. Klaus Diepold 2019 Github Repo In this project, we presented a system for predicting opening weekend box office sales from movie trailers by combining metadata (e.g., actors, studio, genre) with audio-visual features (e.g., spectrograms, colors, emotions), and implemented a frontend that enables users to run predictions on arbitrary YouTube trailers.

Generative Modelling using Capsule Generative Adversarial Networks Bachelor Thesis supervised by Alexander Kuhn, Prof. Alois Knoll, and Prof. Klaus Diepold 2018 Github Repo This thesis introduces CapsGAN, a generative modeling framework that integrates capsule networks into the GAN discriminator to better preserve spatial relationships, evaluates different CapsGAN architectures and routing algorithms, and demonstrates that dynamic routing improves training stability and sample quality compared to standard GAN approaches.

Computed Tomography Image Reconstruction Project in the scope of Ferienakademie by TUM, University of Erlangen and University of Stuttgart 2018 Github Repo In this project, we develop a deep learning–based learnable filter to enhance the CT image reconstruction pipeline for improved image quality and reconstruction accuracy.