SCANNET
PulseAugur coverage of SCANNET — every cluster mentioning SCANNET across labs, papers, and developer communities, ranked by signal.
5 day(s) with sentiment data
-
New framework GMPCR improves multiview point cloud registration in low-overlap scenes
Researchers have developed GMPCR, a novel framework for multiview point cloud registration, particularly effective in scenes with limited overlap. This non-learning-based approach constructs a refined compatibility stru…
-
New research explores semantic retention in 3D segmentation model composition
Researchers have investigated how semantic information is retained when combining frozen foundation models for few-shot 3D segmentation. Their study, titled "Beyond Argmax: A Mechanistic Study of Semantic Retention in F…
-
TUM researchers use 'inverse self-supervision' to advance 3D spatial AI
Researchers from the Technical University of Munich, led by Professor Angela Dai, are developing a novel approach to 3D spatial intelligence that addresses the limitations of current AI models in understanding real-worl…
-
GoDeep uses language-space lifting for annotation-free 3D scene understanding
Researchers have developed GoDeep, a novel method for annotation-free open-vocabulary 3D scene understanding. Unlike typical approaches that embed CLIP features into 3D, GoDeep utilizes a vision-language model purely as…
-
New MSSP framework advances unsupervised 3D point cloud segmentation
Researchers have developed a new framework called MSSP for unsupervised semantic segmentation of 3D point clouds. This method combines multi-scale spectral analysis with spatially-constrained clustering to capture hiera…
-
Scal3R method enhances online 3D reconstruction with multi-relative pose querying
Researchers have developed Scal3R, a novel approach to improve online 3D reconstruction for long videos. By reformulating the problem as multi-reference relative pose querying using lightweight tokens and a frozen backb…
-
VOIM system generates 3D instance maps without training
Researchers have introduced VOIM, a novel training-free system for creating 3D instance maps from RGB-D or monocular RGB data. Unlike existing online systems that commit to labels early, VOIM defers these decisions unti…
-
3D-MRL introduces nested multimodal 3D representations for varied computational budgets
Researchers have introduced 3D Matryoshka Representation Learning (3D-MRL), a novel framework for pre-training multimodal 3D representations. This approach allows a single model to generate embeddings at various dimensi…
-
New Ex-Sim(3)-Reg method enhances 2D-3D registration accuracy
Researchers have developed a new method called Ex-Sim(3)-Reg to improve the accuracy of 2D-3D correspondence pruning in image-to-point-cloud registration. This technique addresses limitations in existing methods that st…
-
New camera pose refinement method bypasses triangulation errors
Researchers have developed a new method for camera pose refinement that avoids triangulation, a common step that can introduce errors. This triangulation-free approach, which expresses structure as scalar depths, mainta…
-
New benchmark OV3D-Bench highlights semantic challenges in 3D detection
Researchers have introduced OV3D-Bench, a new benchmark designed to evaluate open-vocabulary monocular 3D detectors under more realistic deployment conditions. The benchmark addresses inconsistencies in existing evaluat…
-
STAR framework enhances 3D scene understanding with novel routing techniques
Researchers have developed STAR, a novel framework designed to improve 3D scene understanding by addressing challenges posed by topological discrepancies across different sensor modalities. STAR utilizes a Mixture-of-Ex…
-
New PinpointQA benchmark tests MLLMs on indoor video spatial understanding
Researchers have introduced PinpointQA, a new benchmark designed to evaluate the spatial understanding capabilities of multimodal large language models (MLLMs) when processing indoor videos. The benchmark, built upon Sc…
-
New Self-Geometry pipeline enhances 3D vision model geometric consistency
Researchers have developed Self-Geometry, a novel test-time adaptation pipeline designed to enhance the geometric consistency of 3D vision foundation models. This method directly imposes explicit multi-view geometric co…
-
HKUST(GZ) unveils training-free 3D mapping system for robots
Researchers from The Hong Kong University of Science and Technology (Guangzhou) and Mohamed bin Zayed University of Artificial Intelligence have developed FreeOcc, a novel system for semantic occupancy prediction in 3D …
-
LLM Agents Generate Cinematic Camera Paths for 3D Scenes
Researchers have developed CinemaTraj, a novel framework that leverages LLM agents to generate cinematic camera trajectories for 3D scenes. This system takes a set of RGB-D images and a natural language prompt to decomp…
-
GeoStereo framework unifies stereo geometry estimation with diffusion priors
Researchers have introduced GeoStereo, a novel framework that unifies stereo geometry estimation for both disparity and surface normal prediction. This approach leverages diffusion priors to enhance performance in chall…
-
New method enhances 3D scene generation using latent space flow matching
Researchers have developed a new method called Latent Riemannian Flow Matching to improve 3D scene generation using geometric foundation models. This technique operates within the latent space of models like the Visual …
-
New methods advance panoptic segmentation with view synthesis and occlusion awareness · 2 sources tracked
Researchers are developing new methods for panoptic segmentation, a task that involves identifying and delineating every object instance and semantic region within an image. One approach leverages large view synthesis m…
-
New AI methods generate 3D scenes from multi-view images and sketches
Researchers have developed new methods for generating 3D scene assets from limited visual input. One approach, Scene-SAM3D, extends existing single-image 3D generation models to work with multiple calibrated views, impr…