Multimedia Forensics – Detecting deepfakes in images, video and audio, tracing which generator produced a synthetic image, using vision-language models as detectors, and fighting spoofing and fraud.

Adversarial & Robustness – Attacking and defending forensic detectors: data poisoning and backdoor attacks, robustness testing, and watermarking to certify the authenticity of AI-generated content.

Explainable AI – Opening the black box: mechanistic interpretability, explaining why a deepfake detector makes a decision, and interpreting what depth-estimation networks learn.

Efficient Architectures – Optimizing Vision Transformers, compressing networks with pruning, quantization and distillation, and studying the trade-off between speed and accuracy.

Real-world Applications – Bringing efficient vision to real domains, such as precision sericulture (silkworm rearing), traffic and urban mobility monitoring, and datasets and simulation for smart farming.

Embodied AI – Perception for agents in the physical world, such as estimating pedestrian intentions and learning to predict how scenes evolve over time.

VR / AR – Lightweight 3D object reconstruction, and point clouds and meshes for immersive media.

Medical Imaging – AI for dental radiographs: tooth segmentation, detection of periapical lesions and periodontal disease assessment. Also generating synthetic medical data with diffusion models, and explaining model decisions to clinicians.

Earth Observation – Matching ground-level photos to aerial views, semantic segmentation and change detection on satellite imagery, diffusion models for Earth observation, and using satellite data to verify misinformation.

Core Vision Tasks – The fundamentals of seeing: monocular depth estimation, optical flow, video generation and 3D scene reconstruction.

Edge AI & TinyML – Designing CNNs for embedded and IoT hardware, on-device vision (for example, beehive monitoring), and energy-aware inference.

Offloading & Scheduling – Deciding where and when computation runs: reinforcement-learning-based task scheduling, container autoscaling, and resource allocation across fog and edge infrastructure.

Human–Computer Interaction – Intelligent XR interfaces that support human decisions in safety-critical settings.