Computer Vision

MeshXL: Autoregressively generating high-quality 3D meshes using a novel Neural Coordinate Field (NeurCF) representation and large language model approaches.

New scalable method measures per-unit interpretability in vision DNNs without human evaluation, revealing anti-correlation between model performance and interpretability.

New method efficiently measures how well AI models memorize training data, revealing that open-source models memorize less than expected.

MC-DiT: A novel training paradigm for masked diffusion models achieving state-of-the-art image generation by leveraging clean-to-clean reconstruction.

MaskFactory generates high-quality synthetic data for dichotomous image segmentation, improving model training efficiency and accuracy.

Masked Pre-training empowers a universal, fast zero-shot image denoiser!

ManiPose: Manifold-constrained multi-hypothesis model solves 3D human pose estimation’s depth ambiguity, outperforming state-of-the-art models in pose consistency.

MambaSCI: Efficient, novel deep learning model reconstructs high-quality quad-Bayer video from compressed snapshots, surpassing existing methods.

MambaLLIE: a novel implicit Retinex-aware low-light enhancer using a global-then-local state space, significantly outperforms existing CNN and Transformer-based methods.

MambaAD: Linear-complexity multi-class unsupervised anomaly detection using a novel Mamba-based decoder with Locality-Enhanced State Space modules.

LuSh-NeRF: A novel model reconstructs sharp, bright NeRFs from hand-held low-light photos by sequentially modeling and removing noise and blur, outperforming existing methods.

LRM-Zero: Training large reconstruction models solely on synthetic data, achieving quality comparable to real-data trained models.

LP-3DGS learns to optimally prune 3D Gaussian splatting, achieving significant efficiency gains without compromising rendering quality via a trainable binary mask and the Gumbel-Sigmoid method.

LookHere: Vision Transformers excel at high-resolution image classification by using 2D attention masks to direct attention heads, improving generalization and extrapolation.

AdaptOD: a novel approach for robust OOD detection in long-tailed recognition, dynamically adapting outlier distributions to true OOD distributions using a dual-normalized energy loss for improved acc…

Dynamic Rebalancing Contrastive Learning with Dual Reconstruction (2DRCL) pre-training significantly boosts object detection accuracy, especially for underrepresented classes.

Long-range feedback spiking network (LoRaFB-SNet) surpasses other models in capturing dynamic and static visual cortical representations under movie stimuli, advancing our understanding of visual syst…

LoD-Loc: A novel aerial visual localization method uses lightweight LoD 3D maps & neural wireframe alignment for accurate and efficient 6-DoF pose estimation, surpassing state-of-the-art methods.

LoCo: Memory-efficient location-consistent image features learned via a novel ranking loss, enabling three orders of magnitude memory improvement and outperforming state-of-the-art.

CATOD framework improves text-to-image generation by actively learning high-quality training data to accurately depict out-of-distribution concepts.