The first replay-based continual imitation learning framework that leverages the generative capabilities of World Action Models to synthesize replay trajectories without storing real demonstrations from previous tasks.

TIMEPROVE is a hybrid long-video reasoning framework that generates action-grounded hypotheses efficiently and verifies only sparse RGB evidence using an expensive VLM.

UNIEGO is a unified egocentric encoder trained via hierarchical distillation across ego-exo viewpoints, modalities, and foundation models, using representation-specific proxies as structured mediators.

UniLACT is a Vision-Language-Action model that incorporates geometric structure through depth-aware latent pretraining, enabling downstream robot policies to inherit stronger spatial priors.

We introduce Vision Contextualized Probing (VisCoP), which augments the VLM's vision encoder with a compact set of learnable visual probes that enables efficient domain-specific adaptation with minimal modification to pretrained parameters.

We leverage the complementary nature of egocentric views to enhance LVLM’s understanding of exocentric ADL videos through online ego2exo distillation.

3D FaceShell is a privacy-preserving framework that manipulates how vision-language models interpret 3D face avatars while preserving facial identity and visual realism.

DnA introduces denoising attention to improve discriminability in visual tasks by separating relevant and irrelevant features into distinct interacting subspaces.

MS-Temba is the first Mamba based architecture for action detection in long untrimmed videos that can be trained/tested on NVIDIA Jetson Nano.

TICON is an Omni Tile Contextualizer that can contextualize embeddings from any tile encoder.

CORA is a semi-supervised reasoning segmentation framework that jointly learns from limited labeled data and a large corpus of unlabeled images.

Gigapixel Vision-Concept Knowledge Contrastive pretraining (GECKO) aligns WSIs with a Concept Prior for delivering clinically meaningful interpretability.

A novel generative masked model for hand mesh recovery that synthesizes plausible 3D hand meshes.

LLAVIDAL, a Large Language Vision Model, incorporates 3D poses and relevant object trajectories to understand the intricate spatiotemporal relationships within ADLs.

Ski-models introduce 3D skeletons into the vision-language embedding space to enable effective zeroshot learning for ADL.

A generative framework that reformulates monocular HMR as an image-conditioned generative task.

A frequency-aware attention module to unweave skeleton frequency representations for action recognition.

A novel plug-in module designed for existing semi-supervised segmentation frameworks that offers patch-level supervision.

A novel text-to-motion generation framework. BAMM captures rich and bidirectional dependencies among motion tokens.

We introduce the first Pose Induced Video Transformer: PI-ViT (or π-ViT), a novel approach that augments the RGB representations learned by video transformers with 2D and 3D pose information.

Self-Interpretable MIL (SI-MIL), the first interpretable-by-design MIL method for gigapixel WSIs, which provides de novo feature-level interpretations grounded on pathological insights for a WSI.

We present MAVREC, a video dataset where we record synchronized scenes from different perspectives -- ground camera and drone-mounted camera.

A diversity-inducing pretraining technique, tailored to enhance representation learning in digital pathology.

This paper shows that jointly optimizing ViTs for the primary task and a Self-Supervised Auxiliary Task is surprisingly beneficial when the amount of training data is limited.

This paper explains how to utilize OpenAI's CLIP for long-term action detection in videos.

This paper focuses on designing video augmentation for self-supervised learning, we propose CMMC to make use of other modalities in videos for data mixing.

A framework for learning self-supervised video representation that is invariant to unseen camera viewpoints.

The impacts of the existing self-supervised losses with Joint Learning framework for RL is limited, while there is no golden method that can dominate all tasks.

3DTRL is a light-weighted, plug-and play layer that recovers 3D information of visual tokens and leverages it for learning viewpoint-agnostic representations.

TSU is a new untrimmed daily-living dataset consisting of 51 activities performed in a spontaneous manner, captured from non-optimal viewpoints.

A ConvTransformer network that explores global and local temporal relations at multiple resolutions.

VPN++ is an extension of our VPN model (ECCV 2020). VPN++ hallucinates pose driven features while not requiring costly 3D Poses at inference.