Student Lior Yaacov Seminar Title: Enhancing Object Boundary Depth Estimation: The Impact of Synthetic Fractals on Depth Perception Student: Mohamad Saadi Seminar Title: Model-Based Aquaculture Control: Integrating Bioenergetics Modeling and Extended Kalman Filtering for Accurate State Estimation and Real-Time Optimization.

Student Name: Lior Yaacov Degree: MSc. Electrical Engineering Supervisor: Dr. Igal Bilik Seminar Title: Enhancing Object Boundary Depth Estimation: The Impact of Synthetic Fractals on Depth Perception Seminar Summary: Monocular depth estimation (MDE) remains a challenging task due to the inherent ambiguity of inferring depth from a single 2D image. While recent advances in deep learning have significantly improved MDE performance, existing models often struggle to capture fine-grained details, particularly at object boundaries. In this work, we propose a novel approach that enhances depth estimation by incorporating synthetic fractal-based RGB-D data into the training process. Our method utilizes fractal geometry to generate structured depth information for refining the state-of-the-art model MiDaS. Fractals are complex geometric structures characterized by self-similarity across scales, commonly observed in natural phenomena such as coastlines, mountain ranges, and tree branching. Their recursive structure and spatial complexity make them well-suited for modeling diverse depth patterns. In this work, we fine-tune MiDaS checkpoints using our fractal dataset, demonstrating that even a small number of additional training epochs can lead to meaningful improvements in depth estimation. We evaluate our approach on standard benchmarks, including NYU Depth v2, KITTI and DIML, demonstrating improvements in key depth estimation metrics such as RMSE, AbsRel, and delta accuracy thresholds. In addition, we introduce edge-sensitive evaluation metrics to quantify improvements at object boundaries, showing that fractal-based depth augmentation enhances fine-grained depth precision. Our findings suggest that fractal-based fine-tuning can effectively improve generalization and edge accuracy, offering a lightweight and complementary enhancement to existing MDE models. Student: Mohamad Saadi. Supervisors: Prof. Ofer Hadar, Dr. Alaa Jamal. Faculty: Electrical and computer engineering. Degree: Masters (MSc). Abstract Modern aquaculture demands reliable monitoring and intelligent control strategies to ensure efficient production, reduced environmental impact, and optimized energy use. This study presents a model-based framework that integrates bioenergetics modeling with real-time data assimilation and optimization. A comprehensive dynamic model was developed to describe fish growth, heat loss, electricity consumption, and ammonia dynamics, initialized using a standard parameter set. To adapt the model to real conditions, an Extended Kalman Filter (EKF) was implemented by discretizing the model equations and numerically computing the Jacobian matrices required for state–parameter updates. Data were collected from controlled experiments in the Volcani Institute, and real-time calibration using the EKF on these data significantly improved state estimation—particularly for fish weight—and enhanced predictive reliability by jointly updating both system states and key parameters. The calibrated model was subsequently deployed within a real-time simulation and optimization framework, demonstrating its potential for guiding operational decisions. Overall, the results highlight the value of combining mechanistic models with data assimilation to improve monitoring, forecasting, and optimal control of aquaculture systems.