Eric Marchand’s research while affiliated with French National Centre for Scientific Research and other places

... For real-world deployment, a consistent and accurate estimation module is essential. Dynamic SLAM frameworks [7]- [9] provide a promising solution by converting raw sensor inputs into precise pose and motion estimations, forming the foundation for trajectory forecasting. Unlike the dataset provided ground-truth data that undergoes extensive preprocessing within the dataset as well as often before each prediction model individual training, accurate dynamic SLAM outputs are generated directly from sensor measurements, reducing latency and avoiding the errors introduced during pre-processing. ...

Reference:

Training Trajectory Predictors Without Ground-Truth Data

... Of particular relevance to our work are [23,12,24,28]. These approaches also consider spherical safe space representations, computed by keeping track of obstacle points [12,23,28] or a distance transformation of occupancy grid [24], coupled with specialized sampling strategies. ...

Reference:

Safe Bubble Cover for Motion Planning on Distance Fields

... This offers significant advantages in cinematography: (1) enabling smooth camera movement without requiring gimbals or dollies; (2) supporting virtual re-shoots, thus eliminating the need for expensive in-person re-shoots; (3) producing slow-motion effects without requiring special equipment; (4) altering scene geometry and aesthetics without manual intervention, reducing the burden on visual effects and compositing artists He et al. 2024;Kuang et al. 2023;Song et al. 2023b;Yuan et al. 2022]; and (5) introducing/correcting camera-based artefacts like de-blurring [Wang et al. 2023d] and lens distortions Moreau et al. 2022;Xian et al. 2023]. More creative uses have also been explored, like cinematic style transfer [Wang et al. 2023a], the discovery of new cinematographic shots [Skartados et al. 2024] and modeling live sport for action-replay [Lewin et al. 2023]. In Figure 1, we demonstrate this by training a static NeRF toy scene using only 36 images to synthesize a fly-through shot that would otherwise require special equipment. ...

Reference:

Exploring Dynamic Novel View Synthesis Technologies for Cinematography

... While effective in controlled settings, these classical methods show limited robustness to image perturbations and typically require exact target object instances. Recent learning-based approaches attempt to address these limitations through different strategies such as pose regression [9], velocity regression [10], learned feature extraction [11], metric learning [12], and unsupervised feature learning [13], [14], [5], [15]. However, these methods introduce new challenges -they require task-specific training, extensive data generation, or predefined object models, making them difficult to deploy in real-world scenarios where unseen objects and environmental changes are common. ...

Reference:

ViT-VS: On the Applicability of Pretrained Vision Transformer Features for Generalizable Visual Servoing

... SphereMap [16] builds a roadmap for UAV exploration in subterranean environments by sampling spherical segments. A roadmap for exploration of mobile robots can be built by sampling disks and connecting them with a visibility graph [17]. We do not consider these methods in our evaluation, because they are optimized for area coverage of a sensor or for exploration, not for path planning. ...

Reference:

GSRM: Building Roadmaps for Query-Efficient and Near-Optimal Path Planning Using a Reaction Diffusion System

... Utilizing semantic data is another topic that attracted great attention in recent years. Gonzalez et al. [17] presented S 3 LAM, a monocular approach based on ORB-SLAM 2.0 [18] that uses semantic segmentation of generic objects and structures. Their approach fits a plane using the triangulated 3D points for each a priori planar cluster and uses a CNN-based panoptic segmentation framework titled Detectron2 [19]. ...

Reference:

Marker-Based Visual SLAM Leveraging Hierarchical Representations

... This system can be seen as a pinhole camera in a pan-tilt configuration subject to external motions induced by satellite orbit. Pan-tilt motion control using classical VS [30], [12], [18] may lead to motion-blurred images (see Fig. 1). We propose in this paper to constrain the camera rotational velocity using a Quadratic Programming (QP) scheme for limiting motion blur as much as possible while the satellite performs its task. ...

Reference:

QP-based Visual Servoing Under Motion Blur-Free Constraint

... Qiu et al. introduced rigid and motion constraints to model articulated objects on the basis of dynamic rigid object tracking, achieving self-positioning, rigid object tracking and articulated object tracking [30]. The constrained SLAM in dynamic environment (TwistSLAM) proposed by Gonzalez et al. created point clusters based on the semantic classes of dynamic objects [31]. Jointly estimated the pose between moving objects and cameras by constructing inter-cluster constraints of semantic classes. ...

Reference:

Real-Time Dynamic Visual-Inertial SLAM and Object Tracking Based on Lightweight Deep Feature Extraction Matching

... In contrast, the second method minimizes the distance between a data point and the corresponding model point plane, which is perpendicular to its normal plane. Generally, distance point-to-plane minimization converges faster and with greater accuracy, but not necessarily a better region of convergence [11]. ...

Reference:

An improved iterative closest point algorithm based on the particle filter and K-means clustering for fine model matching

... On the other hand, some researchers have tried to obtain more robust image features [11][12][13] for VSLAM to overcome the adverse effects caused by insufficient illumination. Further, some researchers [12,14,15] have extensively researched low-illumination VSLAM and explored the effective combination of image enhancement algorithms and VSLAM systems. However, most low-light VSLAM systems are designed for low-light environments with artificial light, and there is a lack of research on dark night scenes without artificial light. ...

Reference:

Dark-SLAM: A Robust Visual Simultaneous Localization and Mapping Pipeline for an Unmanned Driving Vehicle in a Dark Night Environment