Article

After feedforward reconstruction becomes 1,200 times faster, is BA still necessary in dynamic 4D reconstruction?

StreamSplat at ICLR 2026 achieves purely feedforward, online generation of dynamic 3DGS directly from uncalibrated video streams, at 1,200 times the speed of optimization-based methods. But MoRe at CVPR 2026 actively embeds a BA-like refinement module in its purely feedforward pipeline; removing it raises Sintel ATE from 0.147 to 0.155. BA-Track, an ICCV 2025 Oral, uses a 3D point tracker to decompose camera and object motion, adapting classic BA to dynamic scenes.

Part of the event: After feedforward reconstruction becomes 1200× faster, dynamic 4D reconstruction still needs BA

This page provides a summary and source link only.