The field of computer vision is moving towards more accurate and robust methods for dynamic scene reconstruction and camera motion understanding. Researchers are exploring novel approaches to separate camera-induced motion from the observed motion of dynamic objects, enabling more reliable bundle adjustment and depth refinement. Another area of focus is the development of methods that can infer relative depth from monocular videos by examining the spatial relationships and temporal evolution of tracked 2D trajectories. Noteworthy papers include:

• Back on Track: Bundle Adjustment for Dynamic Scene Reconstruction, which leverages a 3D point tracker to operate reliably on all scene elements.
• Seurat: From Moving Points to Depth, which proposes a novel method that infers relative depth by examining the spatial relationships and temporal evolution of a set of tracked 2D trajectories.
• TAPIP3D: Tracking Any Point in Persistent 3D Geometry, which represents videos as camera-stabilized spatio-temporal feature clouds to enable robust tracking over extended periods.
• Towards Understanding Camera Motions in Any Video, which introduces a large-scale dataset and benchmark designed to assess and improve camera motion understanding.
• Dynamic Camera Poses and Where to Find Them, which introduces a large-scale dataset of dynamic Internet videos annotated with camera poses.