All papers

This paper presents a flow-matching transformer pipeline to generate full gigapixel histopathology whole-slide images.

This paper introduces a method for multimodal models to reason over images using token-level latent regions instead of bounding-box coordinates.

This paper presents a test-time diffusion method that suppresses text hallucinations during image super-resolution by selectively generating text.

This paper develops an immersion-based recurrent network approach for trajectory tracking and navigation on curved geometric manifolds.

This paper benchmarks a pipeline combining real-time object detection with recurrent neural networks to forecast short-term traffic density.

This paper proposes pretraining world-action models by predicting future actions and frames across paired human demonstration videos.

This paper introduces a framework that acquires and directly executes reusable visual programs to improve multimodal reasoning agent reliability.

This paper studies transformer architectural choices for handling missing modalities and longitudinal observations in multimodal healthcare datasets.

This paper introduces a parameter-efficient diffusion method to composite video foregrounds into unaligned background reference scenes.

This paper proposes a cross-modal representation fusion method to predict long-term tool trajectories during microsurgical anastomosis procedures.

This paper proposes a framework that learns dynamic skeletal joint relationships and temporal interactions for 3D human pose estimation.

This paper evaluates spatial and temporal modeling methods to improve multi-class tissue recognition during surgical optical trocar insertion.

This paper presents a unified model for dense visual correspondence across 25 pairs of multi-modal sensor inputs.

This paper analyzes intermediate layer activations in diffusion transformer backbones to explain why layer performance follows an inverted-U curve.

Proposes a world model combining dense bird's-eye-view representations and motion queries for roadside dynamic occupancy forecasting.

Introduces an adversarial training method to make few-step autoregressive video models robust against accumulated rollout errors.

Proposes a training-free KV cache consolidation technique guided by historical queries to manage memory in long video generation.

Develops a context-scheduling method that preserves anchor frames and evicts stale history in long-horizon robotic world models.

Introduces a training-time sampling method to prevent action proposal collapse in vision-language-action models tested with verifiers.

Presents a dataset and vision-language alignment framework for fine-grained human gait attribute recognition from walking sequences.

Combines wheel kinematics with Time-of-Flight camera geometry using directional covariance shaping to prevent odometry divergence.

Introduces a benchmark and multimodal agent that uses visual inspection and external retrieval to verify real-world image factuality.

Develops a dedicated flow estimator and reference warping method for dual-lens super-resolution under large parallax conditions.

Presents an attention-based assembly feature mate prediction model to automate CAD assembly modeling from unassembled boundary representation parts.

Investigates whether synthetic aerial imagery benchmarks preserve model-selection rankings compared to evaluations on real drone data.

Proposes a text anchor re-grounding method using low-rank adapters to improve concept recognition in Segment Anything Model 3.

Accelerates few-step autoregressive video diffusion by predicting features from historical sampling trajectories across video chunks.

Introduces an adaptive quantization and packing framework to compress 4D Gaussian Splatting dynamic scenes using standard video codecs.

Presents an adaptive retrieval framework that selectively retrieves historical video context to assist streaming question answering.

Develops a training-free sparse attention routing method that adapts competition domains across attention heads in video diffusion transformers.

Proposes a causal joint attention model across faces and time for low-latency active speaker detection in multi-person video.

Introduces a benchmark probing physical perception tasks related to humans, including depth, trajectory, illumination, and kinematics.

Presents an autoregressive video rerendering pipeline using a dynamic 4D point-cloud memory cache for novel-view video synthesis.

Refines feed-forward 3D reconstruction models on dynamic video through unlabeled post-training using relative depth constraints and 2D priors.

Integrates real-world image retrieval into diffusion models using pixel-level affinity maps and heterogeneous noise initialization to accelerate generation.

Proposes a video autoencoder compression framework that uses temporal low-rank approximation to reduce computation in latent video diffusion.

Introduces modifications to set-based variational autoencoders to improve geometric alignment and attention quality in 3D shape reconstruction.

Applies Direct Preference Optimization to align generative models with human handwriting preferences and capture natural shape variability.

The paper presents an architecture that decouples video and action prediction models using latent bridge queries.

The paper presents a reproducibility study and an explicit saliency supervision extension for adversarial fine-tuning of CLIP vision models.

The paper presents an indoor paired thermal and RGB dataset and benchmarks for privacy-preserving human occupancy detection.

The paper proposes a motion capture framework that predicts and aligns joint rotations across arbitrary target skeleton topologies from monocular video.

The paper introduces a framework for continuous oblique thermal drone geo-localization using 2D satellite maps in navigation-denied settings.

The paper proposes learning motion-oriented register tokens during video pretraining to provide compact state representations for driving world models.

The paper introduces a hierarchical commander-executor framework to coordinate multi-agent cooperative object manipulation tasks in physics simulations.

The paper proposes an engram-neuron-inspired memory framework using gradient representations to regulate updates in lifelong learning.

The paper presents a parameter-efficient architecture connecting text embeddings with mask decoding for language-guided medical image segmentation.

The paper proposes an analytical projection method for drone visual localization over 2D maps without digital surface models.

The paper presents a framework that uses structured 3D facial guidance as in-context input conditioning for controllable portrait video animation.

The paper proposes an anatomical density-guided 3D Gaussian splatting method for sparse-view dental computed tomography reconstruction.

The paper proposes a reinforcement learning framework using natural language metrics to reduce object hallucination in vision-language image captioning.

The paper evaluates layer input-output cosine similarity to prune redundant layers in vision transformers without retraining.

The paper presents a visual tracking and analysis framework for multi-camera cattle monitoring and heat-stress detection.

The paper theoretically analyzes whether volume rendering with independent sources can reproduce external views while maintaining incorrect internal density.

The paper conducts a controlled benchmark study evaluating common contrastive training strategies for universal multimodal embedding learning.

The paper compares five convolutional neural network architectures for detecting branch dieback disease in pequi tree images.

The paper analyzes historical World Risk Index data using Python visual analytics libraries to assess global natural disaster vulnerabilities.

The paper evaluates lightweight convolutional networks and vision transformers for multi-class brain tumor classification from MRI scans.

The paper introduces spatially guided pooling modules to enable region-aware channel recalibration in convolutional networks for image segmentation.

The paper introduces a framework that distills iterative prompt search into rule modules for efficient text-to-video generation.