Computer Vision

Researchers developed SF-V, a single-step image-to-video generation model, achieving a 23x speedup compared to existing models without sacrificing quality, paving the way for real-time video synthesis…

HERT: a novel framework for semi-open 3D object retrieval using hierarchical hypergraph equilibrium, achieving state-of-the-art performance on four new benchmark datasets.

SemFlow: A unified framework uses rectified flow to seamlessly bridge semantic segmentation and image synthesis, achieving competitive results and offering reversible image-mask transformations.

Universal Unsupervised Cross-Domain Retrieval (U2CDR) framework learns semantic features to enable accurate retrieval even when category spaces differ across domains.

Self-Play Fine-Tuning (SPIN-Diffusion) revolutionizes diffusion model training, achieving superior text-to-image results with less data via iterative self-improvement, surpassing supervised and RLHF m…

Self-guided MAE boosts self-supervised learning by intelligently masking image patches based on internal clustering patterns, dramatically accelerating training without external data.

Self-Distilled Depth Refinement (SDDR) tackles noisy depth maps via a novel noisy Poisson fusion approach, achieving significant improvements in depth accuracy and edge quality.

Unsupervised SAM (UnSAM) achieves competitive image segmentation results without human annotation, surpassing previous unsupervised methods and even improving supervised SAM’s accuracy.

AnyChange achieves zero-shot image change detection by adapting the Segment Anything Model (SAM) via a training-free bitemporal latent matching method, significantly outperforming previous state-of-th…

SE(3)-bi-equivariant Transformers (BITR) revolutionizes point cloud assembly by guaranteeing robust alignment even with non-overlapping clouds, thanks to its unique equivariance properties.

SCube: Instant large-scale 3D scene reconstruction from sparse images using VoxSplats, a novel 3D Gaussian splat representation.

Researchers improved 3D shape generation from 2D diffusion models by showing that existing Score Distillation Sampling is a reparameterized version of DDIM and fixing its high-variance noise issue via…

Logistic Schedule: A novel noise schedule revolutionizes image editing by improving DDIM inversion, enhancing content preservation and edit fidelity without model retraining!

CRATE-a: A new white-box vision transformer architecture achieves 85.1% ImageNet accuracy by strategically scaling model size and datasets, outperforming prior white-box models and preserving interpre…

VQGAN-LC massively scales VQGAN’s codebook to 100,000 entries while maintaining a 99% utilization rate, significantly boosting image generation and downstream task performance.

ScaleKD: Pre-trained vision transformers make excellent teachers for diverse student networks, improving efficiency and performance in knowledge distillation.

Samba: a novel severity-aware recurrent model, tackles cross-domain medical image grading by sequentially encoding image patches and recalibrating states using EM, significantly improving accuracy.

This paper introduces SAM-guided masked token prediction, a novel framework for 3D scene understanding that leverages foundation models to significantly improve 3D object detection and semantic segmen…

RTify: A novel framework aligns deep neural networks’ dynamics with human reaction times for improved visual decision-making models.

SGA+ significantly boosts neural image compression by refining latents, offering a flexible, hyperparameter-insensitive approach with improved rate-distortion trade-off.