Towards Quantum Approaches for Object Detection using QUBO Formulations

Object detection algorithms often struggle with highly redundant predictions (bounding boxes), especially when using heuristic algorithms in crowded scenarios. This thesis investigates a completely different approach, by translating the bounding box suppression task into a global mathematical optimization problem. Specifically, the chosen formulation is the Quadratic Unconstrained Binary Optimization (QUBO) model. In this way, the suppression task becomes a maximization problem: the system balances the confidence scores of each box against a set of spatial penalties. These penalties are computed by evaluating the overlap between pairs of bounding boxes, punishing the selection of redundant predictions to find the best overall combination of surviving boxes. Since the QUBO model is mathematically equivalent to the Ising model, this optimization task is directly solvable using quantum hardware. Through an extensive analysis, this work compares the solutions and execution times on the MS COCO dataset across classical solvers (like the Gurobi Optimizer and Simulated Annealing) and the D-Wave Quantum Annealer. Our experimental results demonstrate that quantum hardware can maintain state-of-the-art detection accuracy while overcoming the exponential time complexity of classical solvers on images with a high density of bounding boxes.