Nuscenes
PulseAugur coverage of Nuscenes — every cluster mentioning Nuscenes across labs, papers, and developer communities, ranked by signal.
16 day(s) with sentiment data
-
ForeseeDrive introduces future-token conditioning for generative driving
Researchers have introduced ForeseeDrive, a novel method for anticipatory generative driving that conditions future scene evolution into query-indexed tokens. This approach allows for shaping candidate generation rather…
-
ExQuery framework enhances temporal 3D object detection with structured query priors
Researchers have developed ExQuery, a novel framework for temporal 3D object detection that enhances the encoding of semantic, geometric, and motion information. This method organizes decoded queries into shared semanti…
-
BEVCast: Adaptive BEV System for Efficient Autonomous Driving
Researchers have developed BEVCast, a novel two-stage Bird's-Eye-View (BEV) system for autonomous driving that adapts its resolution based on the predicted driving path. This approach optimizes computational resources b…
-
FlowSim bypasses image rendering for faster end-to-end driving simulation
Researchers have developed FlowSim, a novel simulation method for evaluating end-to-end driving systems that bypasses the need for image rendering. This approach directly predicts internal latent representations, signif…
-
Video2Asset framework constructs persistent 3D street-scene assets from video
Researchers have introduced Video2Asset, a novel online framework designed to construct persistent 3D street-scene assets from multi-view video. This system addresses the challenge of accumulating intermittent observati…
-
FreeOcc model efficiently forecasts 4D occupancy by decoupling ego-motion
Researchers have developed FreeOcc, a novel 4D occupancy forecasting model that efficiently disentangles ego-motion from scene dynamics. This model utilizes a continuous-time framework with explicit SE(2) geometric deco…
-
New WAM-Sim model enhances driving simulation realism with hierarchical memory
Researchers have developed WAM-Sim, a novel driving world model designed to overcome the limitations of autoregressive models, which often suffer from structural and identity drift over long simulation horizons. WAM-Sim…
-
New framework integrates VLM reasoning into 3D object detection without added cost
Researchers have developed MonoLaDN, a novel framework for monocular 3D object detection that leverages the geometric reasoning capabilities of vision-language models (VLMs) without increasing inference costs. During tr…
-
OccSplat framework enhances 3D driving scene reconstruction by addressing occluded geometry
Researchers have developed OccSplat, a novel feed-forward framework designed to improve driving scene reconstruction from surround-view images. Unlike previous methods that struggle with occluded regions, OccSplat expli…
-
MoSoLoco: New Monocular 3D Human Localization System for Robots
Researchers have developed MoSoLoco, a novel monocular 3D human localization system designed for resource-constrained robotic hardware. This system leverages internal visual features from pose estimators along with dete…
-
CalibSwitch framework enhances camera-LiDAR sensor calibration for multimodal perception
Researchers have developed CalibSwitch, a novel framework designed to improve the accuracy and reliability of camera-LiDAR sensor calibration for multimodal perception systems. The system separates the process into abso…
-
New Object-Aware Model Enhances Autonomous Driving Map Prediction
Researchers have developed ObjectGuidedMap, a novel multitask architecture for predicting Bird's Eye View (BEV) maps in autonomous driving systems. This model jointly predicts BEV maps and 3D objects, allowing for infor…
-
UFMF system uses radar for real-time 4D occupancy forecasting
Researchers have developed UFMF, a novel real-time 4D occupancy forecasting system designed for autonomous driving. This system utilizes an attention-free temporal core combined with radar-guided kinematic priors to acc…
-
LiDAR-based token selection slashes compute for autonomous driving models
Researchers have developed a novel method called Geometry-Aware Token Selection (GATS) for autonomous driving models that significantly reduces computational load by selectively processing image tokens. By utilizing LiD…
-
New INTERACT framework learns interaction-aware dynamics for safer autonomous driving
Researchers have developed INTERACT, a novel three-stage framework designed to enhance autonomous driving safety by learning interaction-aware world dynamics from driving videos. This system models how surrounding traff…
-
FastBEVFusion system enhances real-time 3D object detection for autonomous vehicles
Researchers have developed FastBEVFusion, a novel system designed to improve real-time 3D object detection for autonomous driving by fusing data from LiDAR and cameras. This system projects modality-specific features in…
-
New SR-BEVS method boosts camera-based 3D object detection robustness
Researchers have developed a novel approach called Sparse Recovery Bird's-Eye View Model Soup (SR-BEVS) to enhance the robustness of camera-based 3D object detection models, such as Fast-BEV, against real-world corrupti…
-
New method evaluates ML LiDAR safety targets across conditions
A new paper proposes an uncertainty evaluation method to support the Safety Of The Intended Functionality (SOTIF) analysis for machine learning object detectors used in LiDAR systems. The method aims to determine if per…
-
OmniSpace enhances MLLMs for autonomous vehicles with 2D geometry awareness
Researchers have introduced OmniSpace, a novel plug-and-play paradigm designed to enhance the spatial reasoning capabilities of Multimodal Large Language Models (MLLMs) for autonomous vehicle applications. Unlike existi…
-
New IGLOSS method uses image generation for lidar semantic segmentation
Researchers have developed IGLOSS, a novel method for open-vocabulary semantic segmentation of 3D automotive lidar data. This approach bypasses the typical image-text modality gap by using image generation from text to …