SIGGRAPH method adds physical light control to photos
SFU researchers train a neural renderer with differentiable rendering so users can edit lights like in Blender.
Stories per day · 30 days
9 RSS
Related3d-gaussian-splatting43d-reconstruction4robotics4feed-forward3novel-view-synthesis3benchmark2diffusion2flow-matching2human-mesh-recovery2inference2monocular-depth2nvidia2pose-estimation2self-supervised-learning2semantic-segmentation2video-matting23d13d-generation1
92 stories, newest first · one card per story
Stories tagged #computer-vision
NVIDIA profiles startups using vision AI to sort mixed waste
Four Inception companies run NVIDIA edge GPUs to identify plastics, packaging and textiles on fast conveyor belts.
1
Click3R corrects stereo 3D reconstruction with sparse clicks
A lightweight adapter uses user point correspondences, and the authors say one click can cut reconstruction error.
JoVox jointly models voxel hierarchy with discrete flow matching
Authors report 8x training throughput, 1/15 forward FLOPs, and 11% lower Chamfer distance than TRELLIS.2.
InstantHMR recovers 3D human meshes in real time on phones
A Galaxy S23 NPU produces a 6,890-vertex mesh in 5.5 ms, with 40.0 mm PA-MPJPE on 3DPW.
Occlusion detection from disagreement between two monocular distance cues
Occlusion read from apparent height and monocular metric depth disagreement, no training or labels required.
UniRestorer++ scales image restoration with degradation experts
The paper reports roughly log-linear PSNR gains and a 50-category CDD-50 benchmark.
Adversarial hull patches and physical disguise evaluated for USV detection
Physical disguise bypasses camera and radar checks while adversarial patches suppress detection in up to 29% of attempts.
ReLift recasts dense 3D tracking as trajectory repair
ReLift repairs lifted trajectories from pretrained 2D and 3D models to achieve 2.1 APD gains on WorldTrack without retraining.
Study tests multi-view strategies for real-world luggage re-identification
An end-to-end pipeline detects and re-identifies luggage across video streams despite occlusion, clutter and crowds.
SIFR fuses CLIP and DINOv2 spectrally for domain-general segmentation
The method filters domain-sensitive responses at fusion and stabilizes multi-scale decoder features before dense prediction.
FDSML uses formula parameters for metric learning without real images
On ResNet-50, Recall@1 rises from 65.6% to 74.0% on Stanford Online Products and from 56.2% to 78.4% on Stanford Cars versus FDSL.
RePAIR repairs backdoored semantic segmentation models without retraining
The post-hoc, label-free method infers victim-target pairs from unlabeled images and corrects suspicious pixels.
Paper proposes Quaternion Approximate Networks for vision tasks
The authors report better accuracy with fewer parameters than prior convolution and quaternion models on standard vision benchmarks.
Linear probe predicts VGGT pose failures the model does not report
On HardMatch, where VGGT fails on 35% of 1,000 pairs, a frozen-feature probe reaches 93.5% ROC-AUC.
MatAnyone-E cuts 4K video-matting cost with sparse memory matching
Authors report about 7.4x faster inference and peak GPU memory at 7.6% of MatAnyone 2 at 4K.
OBB-PBD extends parallel box decoding to oriented aerial boxes
Reported rotIoU on the full 6139-record AVVG test set rises from 0.1253 zero-shot to 0.303.
Easy-3Dfy composes object meshes into single-image 3D scenes
An RPF aligner matches object-local meshes to partial scene pointmaps without full scene supervision.
F2G-Pose estimates category-level RGB-D object pose in one pass
It reports state-of-the-art accuracy on SOPE and ROPE at 28 FPS for pose-only inference, with code and weights released.
SRCD distills head-class CAMs to resist shortcuts on tail classes
The method aggregates a few head-class activation maps into a teacher and reports consistent gains on long-tailed benchmarks.
Learned 3DGS edits power online Bayesian scene change detection
Authors report 67.64% mIoU on PASLCD and 75.82% on CCS3D, with the full PASLCD pipeline at 17.2 FPS.
NVIDIA model co-generates video and 3D motion for arbitrary rigs
Cross-attention links a skeleton model to a pretrained video model, trained on nearly 200K pairs.
WildMAE extends masked autoencoders for wildlife camera-trap images
On 4,744 Snapshot Serengeti images at 75% masking, historical PSNR was 27.90 dB for WildMAE versus 24.33 for vanilla MAE.
HRoPE encodes patch surfaces as slanted planes for multi-view attention
Plane parameters need no ground-truth geometry and can be learned from view synthesis alone or taken from foundation models.