Open-Vocabulary Audio-Visual Event Localization via Complex-Valued Fusion
The paper addresses Open-Vocabulary Audio-Visual Event Localization by replacing standard fusion rules with a complex-valued neural network (CVNN). This method treats visual and audio representations as real parts of complex numbers, using companion streams to supply the imaginary parts. The resulting complex convolution yields richer similarity scores, leading to state-of-the-art performance on OV-AVEL benchmarks.