A5021651868- Robotics and Sensor-Based Localization
- Robot Manipulation and Learning
- Advanced Vision and Imaging
- Advanced Image and Video Retrieval Techniques
- Electronic and Structural Properties of Oxides
- Gas Sensing Nanomaterials and Sensors
- Generative Adversarial Networks and Image Synthesis
- Underwater Vehicles and Communication Systems
- Image and Object Detection Techniques
- Catalysis and Oxidation Reactions
- Human Pose and Action Recognition
- Advanced Image Processing Techniques
- Image Processing Techniques and Applications
- Industrial Vision Systems and Defect Detection
École Nationale Supérieure d'Ingénieurs de Caen
2025
Université de Caen Normandie
2025
Centre National de la Recherche Scientifique
2025
Normandie Université
2025
Robert Bosch (Germany)
2022-2024
Universität Ulm
2022
Technical University of Munich
2014
We present the D-eDVS- a combined event-based 3D sensor - and novel full-3D simultaneous localization mapping algorithm which works exclusively with sparse stream of visual data provided by D-eDVS. The D-eDVS is combination established PrimeSense RGB-D biologically inspired embedded dynamic vision sensor. Dynamic sensors only react to contrast changes output in form events represent individual pixel locations. demonstrate how an can be fused classic frame-based produce depth-augmented...
Abstract In the growing field of low‐cost electronics, epitaxy complex oxide thin films on a Si substrate requires significant technical means. Therefore, large attention is paid to release freestanding interest from its deposition support which then placed onto substrate, via etching an intermediate sacrificial layer. The use layer offers several advantages since flexible polymer exploited for transfer can also be fully utilized design heterostructure. For green technology, more and...
We propose a framework for robust and efficient training of Dense Object Nets (DON) [1] with focus on industrial multi-object robot manipulation scenarios. DON is popular approach to obtain dense, view-invariant object descriptors, which can be used multitude downstream tasks in manipulation, such as, pose estimation, state representation control, etc. However, the original work focused singulated objects, limited results instance-specific, applications. Additionally, complex data collection...
We propose a self-supervised training approach for learning view-invariant dense visual descriptors using image augmentations. Unlike existing works, which often require complex datasets, such as registered RGBD sequences, we train on an unordered set of RGB images. This allows from single camera view, e.g., in robotic cell with fix-mounted camera. create synthetic views and pixel correspondences data find our are competitive to the methods, despite simpler recording setup requirements. show...
Robot manipulation relying on learned object-centric descriptors became popular in recent years. Visual can easily describe task objectives, they be efficiently using self-supervision, and encode actuated even non-rigid objects. However, learning robust, view-invariant keypoints a self-supervised approach requires meticulous data collection involving precise calibration expert supervision. In this paper we introduce Cycle-Correspondence Loss (CCL) for dense descriptor learning, which adopts...
We propose a framework for robust and efficient training of Dense Object Nets (DON) with focus on multi-object robot manipulation scenarios. DON is popular approach to obtain dense, view-invariant object descriptors, which can be used multitude downstream tasks in manipulation, such as, pose estimation, state representation control, etc.. However, the original work focused singulated objects, limited results instance-specific, applications. Additionally, complex data collection pipeline,...