Foundation4D links 3D detections into full-recording tracks, tops NuScenes without retraining
lealtaixe · x · 2026-10-29
NVIDIA's Foundation4D detects 3D objects, estimates velocities, and links detections into tracks across an entire recording. Longer temporal context improves detection without retraining, achieving top results on NuScenes.
The same shared features also predict 3D Gaussians for static scenery and moving actors, letting a reconstruction branch render novel viewpoints without per-scene optimization, unifying perception and scene appearance.
More from Embodied
- Raven Prism launches: a full Linux computer on your face, first batch ships in January — Scobleizer · 2026-10-29
- RealSense's new Min-Z update lets robots see objects just 2cm from their fingertips — chrismatthieu · 2026-10-29
- Humanoid robots don't need to scan the whole ground: structured perception could slash terrain training — breadli428 · 2026-10-29
- New Meshtastic tutorial shows how to set up a solar-powered LoRa repeater — ShawnHymel · 2026-10-29
- GPT-6 Sol beats Gemini 3.8 Flash 1012–972 on Physical Agent Capability Index — qinzytech · 2026-10-29
- Mila's open-source Harpia uses drones + AI to survey hundreds of tree crowns in a day — Mila_Quebec · 2026-10-29