Motion Mask-driven Improvements in Monocular Dynamic Novel View Synthesis

Suwoong Yeom, Hosung Son, Chanhee Kang, Joonsoo Kim, Kugjin Yun, Won-Sik Cheong, Suk-Ju Kang

Abstract: Monocular dynamic neural radiance field models have usually evolved through a method that separates dynamic and static objects, optimizing each independently. The networks responsible for rendering static and dynamic objects are separate, and the final visualization is achieved by merging the outputs of both systems. Static and dynamic objects are distinguished by a motion mask. This mask results from segmenting only the moving objects within a scene, and it is essential in monocular dynamic neural radiance field models for distinguishing between static and dynamic objects. monocular dynamic neural radiance field models primarily utilize the Mask R-CNN [1]model, without placing a concentrated emphasis on precisely distinguishing moving objects. We obtained a more precise motion mask using a hand-crafted approach and confirmed that the rendering quality improved compared to existing methods. As a results, we verified that accurate motion mask estimation in dynamic neural radiance fields can be an important factor in enhancing performance.