Paper by NAVER LABS Accepted to ICCV 2019
NAVER LABS’s paper “Drop to Adapt: Learning Discriminative Features for Unsupervised Domain Adaptation” has been accepted to the International Conference on Computer Vision (ICCV) 2019, an international academic conference of the highest prestige in computer vision and pattern recognition.
The authors of this paper are Researcher Nam-il Kim of the NAVER LABS Autonomous Driving Group and two interns. This paper is about a study on domain adaptation to utilize the data from virtual environments as real data for deep-learning. It proposes a methodology that is simpler than existing methods, while delivering stronger performance that can be widely applied to existing image-based models.
Domain adaptation to successfully apply a deep-learning model trained using existing data (e.g. virtual data, camera A) to a new set of data (e.g. actual data, camera B) is a subject of study drawing attention in various fields. In particular, the necessity of this method is being emphasized for various autonomous driving and robotics applications, such as not only when the platform sensor is changed, but also when the virtual simulator is applied in an environment where it is difficult to obtain real data. This paper proposes a method of changing the feature space formed with existing data (source domain) to ensure that new data (target domain) are successfully categorized on the basis of the machine learning theory and mathematical modeling without the tagging information of the target domain in the feature space. The proposed method can be applied to a range of deep-learning models, and it delivered excellent performance regardless of the size of the dataset.
In addition, NAVER LABS Europe made the study result public through a poster session and workshop. It is expected that the results of this paper will be utilized in diverse computer vision-related fields and services.
- Learning with Average Precision: Training Image Retrieval with a Listwise Loss (Jérome REVAUD, Jon Almazan, Rafael SAMPAIO DE REZENDE, Cesar De Souza) Read more >
- Fine-Grained Action Retrieval through Multiple Parts-of-Speech Embeddings (Michael Wray, Diane Larlus, Gabriela Csurka Khedari, Dima Damen) Read more >
- Moulding Humans: Non-parametric 3D Human Shape Estimation from Single Images (Valentin Gabeur, Jean-Sébastien Franco, Xavier Martin, Cordelia Schmid, Grégory Rogez) Read more >
- SLAMANTIC-Leveraging Semantics to Improve VSLAM in Dynamic Environments (Martin Humenberger) Read more >