Overview & Topics
Mainstream computer vision research frequently overlooks factors like on-device latency, resource use, and power consumption. This year's ECV workshop addresses true efficiency in the era of Foundation Models and Embodied AI.
Efficient and Scalable Perception Systems
- Foundation model efficiency, latency-aware training, and tunable accuracy-efficiency tradeoffs
- Scalable architectures (e.g., mixture-of-experts, vision transformers)
- Efficient representation learning for images, multimodal, 3D, and 4D data
- Optimization for edge deployment and low-power hardware
- Annotation-efficient pipelines (simulation, augmentation, active selection)
Generative and Foundation Models on Edge Devices
- Generative AI for language models, parallel decoding, long-context and KV-cache management
- Multimodal generative models, VQA and fusion for text, vision, audio, and sensor data
- Diffusion models, step, adversarial, knowledge distillation, consistency models
- Vision-Language-Action models and interactive environments
3D Scene Understanding and Reconstruction in AR/VR
- 3D from multi-view, sensors, or single images
- Real-time 3D perception, mapping, reconstruction, and rendering
- Compositional and open-vocabulary 3D scene understanding
- Robust segmentation and affordance prediction in complex scenes
- Efficient 3D/4D representation learning and compression (point cloud, 6DoF)
- Multi-view and multi-temporal consistency in 3D/4D generation
Embodied AI and Interactive Agents
- Robotics and embodied vision: active agents, simulation, causal/object-centric representations
- Dynamic modeling of environments and objects for manipulation and navigation (VLA)
- Few-shot and adaptive learning for Vision-Language Agents (VLA) in open environments
- Closed-loop policy learning for task completion in cluttered environments
- Integration of human cues (gaze, pose) into perception and decision-making
- Continual learning, adaptive memory, and long-horizon reasoning
- Evaluation frameworks for perception, action, and memory in embodied systems
Avatars, Neural Rendering, and Personalization
- Expressive avatar synthesis and display for HRI (human-robot interaction)
- Neural rendering for dynamic humans and avatars
- Personalization from limited data (few snapshots, generalizable enrollment)
- Relighting, appearance adaptation, and editing under varied conditions
- Consistency and realism across environments and devices
- Efficient streaming and transmission of animated avatars
Egocentric and Social Perception
- Egocentric hand/body tracking and human-object interaction
- Emotion-aware communication and feedback in interactive systems
- Strategies for social presence and trust in companion robots
- Methods for continual adaptation and personalization in interactive systems
Potential Impacts
Efficient computer vision on mobile, wearable, and robotics devices unlocks transformative possibilities across industries.
-
Healthcare & Fitness
- Remote Monitoring: Wearable devices monitoring vital signs and physical activities.
- Early Diagnosis: Enhanced image processing for detection of skin conditions, eye health, etc.
-
Industrial Applications
- Quality Control: Automated product inspection for manufacturing.
- Maintenance and Repair: Wearable devices guiding technicians through complex repairs.
-
Accessibility
- Assistance for the Visually Impaired: Identifying objects and reading text aloud.
- Gesture Recognition: Enabling touchless interaction for users with physical disabilities.
-
Environmental Monitoring
- Wildlife Tracking: Monitoring animal movements and detecting activities via edge devices.
-
Robotics
- Extended Operational Time: Reduced energy consumption enabling longer battery life.
- Real-Time Processing: Faster visual processing for responsive actions.
- Increased Mobility: Compact processing units allow for smaller robots.
Highlights from the 8th ECV Workshop (CVPR 2025)
