Includes recovering and rendering 3D scenes from images and video: 3D Gaussian splatting and NeRF, feed-forward and large reconstruction models, multi-view and monocular depth, structure-from-motion, SLAM and camera pose estimation, dynamic/4D scene reconstruction, surface and mesh reconstruction, and differentiable or inverse rendering.
Investigates large multimodal models and vision-language learning: MLLM architectures and training, visual instruction tuning, multimodal reasoning and chain-of-thought (including RL for visual reasoning), visual token efficiency, hallucination and grounding, CLIP-style contrastive alignment, visual question answering, document and chart understanding, and multimodal agents.
Covers generative models for visual content: text-to-image and text-to-video diffusion, flow matching and rectified flow, diffusion transformers, autoregressive and tokenizer-based visual generation, unified understanding-generation models, few-step distillation and sampling acceleration, guidance, and reward-based fine-tuning of generative models.
Covers improving and forming images: super-resolution, denoising, deblurring, deraining/dehazing and all-in-one restoration, low-light enhancement and HDR, image and video compression, image signal processing and demosaicing, computational photography and novel sensors (event cameras, SPAD, light field, polarization), inverse problems, and image quality assessment.
Includes core recognition tasks on images: object detection and instance/semantic/panoptic segmentation, open-vocabulary and open-world recognition, promptable and foundation segmentation models (SAM), referring and reasoning segmentation, few-shot and zero-shot recognition, image classification and fine-grained recognition, and small, oriented, camouflaged or salient object detection.
Focuses on perceiving and synthesizing people: human motion generation and text-to-motion, human and hand pose and mesh recovery, 3D avatars and digital humans, human-object and human-scene interaction, face modeling and talking heads, gaze and gesture, egocentric and social understanding of people.
Covers adapting models to new conditions: test-time adaptation and training, domain adaptation and generalization, few-shot and zero-shot transfer with CLIP-style models, prompt and adapter tuning, continual and incremental learning without forgetting, active and semi-supervised learning, federated learning, and robustness to distribution shift and label noise.
Covers understanding dynamic visual content: video large language models and long-video understanding, action and activity recognition, temporal grounding and moment retrieval, video question answering and captioning, streaming and online video perception, video object segmentation and tracking, event and anomaly understanding, and audio-visual learning.
Focuses on vision for agents acting in the physical world: vision-language-action models, robotic manipulation and grasping, imitation and reinforcement learning for visuomotor policies, embodied navigation and instruction following, embodied question answering, world models for planning, simulation and sim-to-real transfer, and humanoid control.
Addresses risks in visual AI: adversarial attacks and defenses, backdoor and poisoning attacks, deepfake and AI-generated-image detection, image forgery localization and forensics, watermarking and copyright protection, membership inference and privacy, machine unlearning and concept erasure, safety and content moderation for generative models, fairness and bias in vision systems, and interpretability for trust.
Applies vision to specialized imaging domains: medical image analysis (radiology, pathology, ophthalmology, surgery, microscopy), medical vision-language models and report generation, remote sensing and earth observation (satellite, aerial, SAR, hyperspectral), agriculture and ecology, scientific imaging and physics-informed vision, and domain-specific foundation models and benchmarks.
Focuses on learning general-purpose visual representations: self-supervised and contrastive learning, masked image modeling, DINO/JEPA-style pretraining, vision foundation models and their scaling, backbone and architecture design, pretraining data and synthetic data for vision, representation analysis and alignment across models and modalities, and transfer to downstream tasks.
Covers the measurement side of computer vision: new benchmarks and large-scale datasets, evaluation metrics and protocols for generation and perception, human preference and aesthetic evaluation, MLLM-based evaluation and judging, benchmark reliability, contamination and saturation, dataset bias and annotation, and capability analyses of vision and multimodal models.
Includes making vision and generative models cheaper: network pruning and quantization, knowledge distillation, token pruning and merging for ViTs and MLLMs, efficient and linear attention, mixture-of-experts and dynamic computation, training-free acceleration and caching for diffusion models, neural architecture search, and on-device or edge deployment.
Focuses on controlling and modifying visual content: instruction-based image and video editing, inpainting and outpainting, subject-driven personalization and customization, style transfer, pose/motion/camera control, image animation, relighting and compositing, concept erasure and customization of generative models.
Covers perception and planning for vehicles and outdoor scenes: end-to-end autonomous driving, bird's-eye-view perception and 3D occupancy, LiDAR/radar/camera fusion and 3D object detection, HD mapping and lane detection, motion forecasting and trajectory prediction, driving world models and simulation, and corner-case and safety evaluation for driving.
Covers synthesizing 3D and 4D content: text- and image-to-3D, mesh and part-aware generation, native 3D generative models and 3D latent diffusion, texture and material generation, scene and layout generation, 4D and dynamic object generation, and procedural or CAD-like structured 3D synthesis.