About us
Espace utilisateur
Education
INSTN offers more than 40 diplomas from operator level to post-graduate degree level. 30% of our students are international students.
Professionnal development
Professionnal development
Find a training course
INSTN delivers off-the-self or tailor-made training courses to support the operational excellence of your talents.
Human capital solutions
At INSTN, we are committed to providing our partners with the best human capital solutions to develop and deliver safe & sustainable projects.
Thesis
Home   /   Thesis   /   Weakly Supervised Methods for 3D Point Cloud Perception

Weakly Supervised Methods for 3D Point Cloud Perception

Abstract

3D point cloud LiDAR techniques have considerably improved over the course of the last few years, reaching industrial-ready status in terms of methods as well as sensors. In the context of autonomous driving and driving assistance, LiDAR data can provide a rich and reliable information about the environment. Indeed, numerous methods proposed in the literature achieve state-of-the-art performances using this data modality alone on detection, tracking or segmentation tasks. These methods are often inspired by models developed for images in the visible spectrum and leveraging properties specific to 3D LiDAR point clouds to enhance the model efficiency and quality.

From a scientific perspective, 3D point clouds provides a versatile environment for experimental research, as their rich data structure lends itself to original implementation of concepts also developed on other data modalities. For instance, set-based approaches or spatial attention methods, which have recently gained traction for image models, also apply quite naturally to point clouds.

Although 3D point cloud perception can be considered a mature research subject, several major scientific challenges remain nonetheless:

· Even the largest public datasets cannot represent the wide variety of driving environment, as demonstrated by a few Transfer Learning experiments, even more so for accident-prone situation for which data is scarse. Before considering larger-scale deployment of this technology, LiDAR-based perception methods must be trained over more diverse data.

· Although robust and informative, 3D LiDAR point cloud still contain ambiguities originating from the lack of resolution at a distance and occlusions.

In parallel to the development of supervised methods, unsupervised and weakly supervised methods on text, images and videos have progressed rapidly over the last few years.

Adapting them to 3D point clouds would help take advantage of large unannotated datasets which are more easily acquired. Among these techniques, temporal consistency-based approaches seem particularly suitable for 3D point cloud since they might be combined with spatial consistency more naturally than what can be done with 2D data.

Combining LiDAR data over time has not yet been investigated thoroughly and might also help to lift occlusion ambiguities by merging multiple points of views in the case of a moving sensor, and also increase point density. Fusing the data from other sensors, in particular cameras, is also expected to bring substantial improvements.

The main goal of this thesis is to improve generalization and robustness properties of 3D point cloud perception models. Two complementary directions of research can help toward this objective:

· Leveraging unsupervised and weakly supervised methods such as the ones developed in other application domains.

· Proposing new approaches that take advantage of 3D point cloud properties, for instance temporal or spatial consistency and the availability of a complementary data modality from camera sensors.

Laboratory

Département Intelligence Ambiante et Systèmes Interactifs (LIST)
Service Intelligence Artificielle pour le Langage et la Vision
Laboratoire Vision et Apprentissage pour l’analyse de scènes
Paris-Saclay
Top pencilenvelopegraduation-hatlicensebookuserusersmap-markercalendar-fullbubblecrossmenuarrow-down