SCANNET
PulseAugur coverage of SCANNET — every cluster mentioning SCANNET across labs, papers, and developer communities, ranked by signal.
10 day(s) with sentiment data
-
New 'See and Detect' framework advances image-driven 3D object detection
Researchers have developed "See and Detect," a novel image-driven framework for open-vocabulary multi-view 3D object detection. This approach utilizes multi-view images as the primary source for generating 3D bounding b…
-
EcoVGGT framework boosts 3D video reconstruction speed and efficiency
Researchers have developed EcoVGGT, a novel framework designed to significantly improve the efficiency and streamability of 3D reconstruction from long videos. This method utilizes a "Pinned Key Set" of evenly spaced fr…
-
New metric challenges standard camera pose evaluation in computer vision
A new paper questions the standard Absolute Trajectory Error (ATE) metric used in computer vision for evaluating camera pose accuracy. Researchers found that a simple image-free constant-velocity line often outperforms …
-
Feature Puzzles framework enables multi-view 3D segmentation with single-image supervision
Researchers have developed a new framework called Feature Puzzles that enables multi-view consistent segmentation using only single-image semantic segmentation supervision. This approach constructs pseudo-labels through…
-
VoxLution generates high-fidelity 3D indoor scenes with transformer-based approach
Researchers have developed VoxLution, a novel transformer-based method for generating high-fidelity 3D indoor scenes. This approach addresses the common issue of overly smooth geometry in existing models by employing a …
-
New research advances 3D Gaussian Splatting for compression and real-world video
Recent research in 3D Gaussian Splatting (3DGS) focuses on improving compression, handling unsynchronized video, and enhancing feed-forward methods. GSCV++ offers lightweight compression for GS sequences by optimizing p…
-
SplatScout enables VLM-controlled part-level segmentation in Gaussian scenes
Researchers have developed SplatScout, a novel query-driven approach for part-level referring segmentation in Gaussian scenes. This method allows a vision-language model (VLM) to actively control the perception process,…
-
SplatScout enables vision-language models to control part-level segmentation in 3D scenes
Researchers have developed SplatScout, a novel query-driven approach for part-level referring segmentation in Gaussian scenes. Unlike previous methods that rely on predetermined object candidates, SplatScout allows a vi…
-
VFM-Recon advances neural reconstruction by aligning foundation models
Researchers have developed VFM-Recon, a novel method for scene-level neural volumetric reconstruction that effectively bridges the gap between vision foundation models (VFMs) and the scale-consistency requirements of vo…
-
New OVM-Occ framework advances open-vocabulary indoor scene geometry prediction
Researchers have introduced OVM-Occ, a novel Gaussian-based framework for open-vocabulary monocular occupancy prediction in indoor scenes. This method utilizes an instance-aware vision-language teacher and a geometry-aw…
-
MVLoMa enhances multi-view local feature matching with geometry context
Researchers have developed MVLoMa, a novel local feature matching system designed to extend pairwise matching to multiple views. This system enhances geometric understanding by injecting cross-view context into local de…
-
New PLOVIS method tackles data scarcity in 3D point cloud segmentation
Researchers have developed a novel framework called PLOVIS to address data scarcity in 3D point cloud semantic segmentation. This method leverages open-vocabulary image segmentation models to generate pseudo-labels, com…
-
LLMs fine-tuned for QFT and video QA, new research released
Researchers have explored fine-tuning large language models (LLMs) for specialized scientific domains, specifically quantum field theory (QFT). One study compared reinforcement learning (RL) and supervised fine-tuning (…
-
New P3-VLM model tackles 3D vision-language tasks with point-based grounding
Researchers have identified a significant issue in 3D vision-language models (VLMs) where the common use of bounding boxes for spatial prompts creates a "semantic leakage," allowing models to perform well without fully …
-
New framework uses uncertainty to improve 3D segmentation accuracy
Researchers have introduced a novel framework called Uncertainty-Gated Adaptive (UGA) designed to improve generalized few-shot 3D point cloud segmentation. This method uses uncertainty to dynamically control adaptation …
-
Hugging Face researchers unveil HOLI-WORLD for egocentric AI learning
Researchers have introduced HOLI-WORLD, an integrated ecosystem designed to automate the creation of large-scale learning resources from egocentric recordings. This system aims to enhance machine understanding of dynami…
-
New metric VC-PQ improves evaluation of multi-view segmentation
Researchers have introduced View-Consistent Panoptic Quality (VC-PQ), a new metric for evaluating multi-view panoptic segmentation models. Unlike the existing PQ^scene metric, which can be skewed by the area of an insta…
-
SpaceCache boosts point cloud segmentation efficiency by reusing features
Researchers have developed SpaceCache, a novel method for improving the efficiency of semantic segmentation on point cloud streams, which are crucial for autonomous driving and robotics. This training-free feature cache…
-
New Bayesian framework enhances online 3D reconstruction for long video sequences
Researchers have developed BAT3R, a novel framework designed to improve online 3D reconstruction for long video sequences. This framework addresses the progressive degradation seen in existing recurrent architectures by…
-
CompactSplat improves 3D scene reconstruction with adaptive Gaussian allocation
Researchers have introduced CompactSplat, a new feedforward framework designed to improve 3D scene reconstruction from sparse views. This method adaptively allocates 3D Gaussians, concentrating them in complex areas whi…