Data for training AI can be co-occurrent, simulated, or annotated. Co-occurrent data (e.g., past and future frames, audio-video pairs, text-image pairs, user histories) provides meaningful, abundant supervision suitable for generative models and recommendation systems. Simulated data offers infinite abundance, but its utility depends on the simulation quality. Often, neither type is available, necessitating annotations, particularly expensive expert annotations crucial for industrial solutions. This thesis focuses on image annotation efficiency, proposing Human-in-the-loop (HITL) learning: AI assists annotators, whose corrections iteratively improve the AI. Active learning (AL) complements HITL by selecting data strategically for annotation. The thesis investigates HITL theoretically and practically, developing tools incorporating these findings. The thesis has four parts: Part 1: Interactive Image Segmentation (IIS) * IIS accelerates mask annotation with minimal user interaction. Key methods include RITM, SAM, and SegNext. Despite rapid literature evolution, this thesis addresses evaluation robustness and domain shift. Addressing domain generalization, the thesis introduces SegWise, an improvement over WeSAM, combining realistic interaction (from RITM) and SegNext's robust architecture. Part 2: Human-in-the-loop Learning * Theoretically demonstrates how AI can reduce annotators' workload even in simple binary annotation tasks by asking optimally chosen group-level questions. * Experimentally applies HITL in object detection, showing significant annotation acceleration by training AI interactively. * Develops a practical web application for semantic segmentation, using a frozen foundational model rather than end-to-end training, further demonstrating HITL effectiveness. Part 3: Applications * SAR Image Segmentation: Interactive method developed for coal stockpile segmentation, improving 3D estimation. * Solar Panel Segmentation: Iterative annotation using Gaussian spectral modeling aids automatic segmentation for solar farm monitoring. * Generic Annotation Tool: Incorporates AL and foundation model features, enabling efficient custom object detector training (applied in medical cell counting with excellent results). * Geology Wall Detection: Post-processing method (dynamic programming and Bayesian updates) enhances neural network segmentation outputs interactively. Part 4: Reviews * Image Segmentation Review: Unifies segmentation methodologies, proposes robust evaluations, significantly enhances SegNext for automatic segmentation, and introduces structured outputs via superpixels. * Probabilistic Regression Review: Highlights simpler loss functions (pinball, cross-entropy) for training probabilistic regressors, achieving competitive performance and offering valuable uncertainty estimates for AL. In conclusion, the thesis extensively explores efficient image annotation through IIS, HITL theory and practice, targeted industrial applications, and insightful methodological reviews.