Real-time reconstruction of dynamic 3D scenes from uncalibrated video streams is important for real-world applications.
StreamSplat is introduced as a fully feed-forward framework for transforming uncalibrated video streams into dynamic 3D Gaussian Splatting representations online.
Key technical innovations include a probabilistic sampling mechanism in the static encoder and a bidirectional deformation field in the dynamic decoder for robust and efficient dynamic modeling.
StreamSplat has shown better performance in reconstruction quality and dynamic scene modeling compared to previous methods, and supports online reconstruction of long video streams.