Techniques for high-fidelity three-dimensional (3D) reconstruction of a dynamic scene as a set of voxels are provided. One technique includes: receiving, by a processor, image data from each of two or more spatially-separated sensors observing the scene from a corresponding two or more vantage points; generating, by the processor, the set of voxels from the image data on a frame-by-frame basis; reconstructing, by the processor, surfaces from the set of voxels to generate low-fidelity mesh data; identifying, by the processor, performers in the scene from the image data; obtaining, by the processor, high-fidelity mesh data corresponding to the identified performers; and merging, by the processor, the low-fidelity mesh data with the high-fidelity mesh data to generate high-fidelity 3D output. The identifying of the performers includes: segmenting, by the processor, the image data into objects; and classifying, by the processor, those of the objects representing the performers.