Three-dimensional particle tracking velocimetry (3D-PTV) technique is widely used to acquire the complicated trajectories of particles and flow fields. It is known that the accuracy of 3D-PTV depends on the mapping function to reconstruct three-dimensional particles locations. The mapping function becomes more complicated if the number of cameras is increased and there is a liquid–vapor interface, which crucially affect the total computation time. In this paper, using a shallow neural network model, we dramatically decrease the computation time with a high accuracy to successfully reconstruct the three-dimensional particle positions, which can be used for real-time particle detection for 3D-PTV. The developed technique is verified by numerical simulations and applied to measure a complex solutal Marangoni flow patterns inside a binary mixture droplet. Graphic abstract