Fig. 3: Moving object detection results.

a Network architecture and the projection function of the Monet for moving object detection. The previous frame, current frame and next frame are used as three inputs of Monet. The previous frame and next frame are shifted d_prev and d_next pixels, and the current frame is fixed in the projection function. b Interference patterns of a representative frame. The objects are boosted when the projection direction is close to the ground-truth moving direction (labelled with red boxes), while the objects with orthogonal moving directions are suppressed (labelled with yellow boxes). c The input sequences, Monet computational outputs and ground-truth of two representative sequences. The stopped objects are labelled using orange boxes, which are successfully suppressed. Object-level precision-recall (PR) curves and the areas under the curves (AUCs) of each sequence are also presented. An electronic CNN is used for comparison. Note that the moving objects are highlighted in the outputs of Monet, showing very high similarity with the GT. d Comparison results between Monet and the existing optoelectronic saliency detection approach⁵¹. Saliency diff. shows the difference of the next-frame and the previous-frame saliency maps. e The input sequence, Monet simulation results, original and bilateral filtered physical experimental results, and ground-truth of a representative frame in the test sequences. The object-level PR curve and AUC are illustrated on the right. The outputs of Monet simulation and optical Monet experiments show a high correlation in the final distribution. The optical output shows acceptable performance loss compared with simulations. With a simple bilateral solver applied in the final results, the speckles are removed for better visualization. Prev previous, Curr current, Seq sequence, Inter interference, Pat pattern, Norm normalized, GT ground-truth, Sim simulation, Ori original, Exp experimental, Diff difference, DP diffractive propagation