NVIDIA introduces a large-scale called Physical AI Smart Spaces (PAISS) for multi-camera 3D perception in indoor environments like warehouses and hospitals. The dataset includes over 280 hours of synchronized video from nearly 1,800 cameras, with automated annotations for 2D/3D bounding boxes, multi-camera identities, and camera calibration. The benchmark extends conventional 2D tracking evaluation to 3D by introducing a 3D version of the Higher Order Tracking Accuracy (HOTA) metric. It also incorporates via Isaac Sim and Cosmos Transfer appearance augmentation, plus a hidden real-world target to test Sim2Real generalization.
Benchmark tests multi-camera 3D tracking across synthetic and real spaces
The dataset combines synchronized indoor video, dense 3D annotations, appearance-transferred imagery and held-out real-world evaluation.
Big Tech
Yuxing Wang · Yizhou Wang · Anqi Li · Shuo Wang · Sameer Satish Pusegaonkar · Haoquan Liang · +10 more
NVIDIA · Santa Clara University
Research Digest··3 min read
In one line
NVIDIA's Physical AI Smart Spaces benchmark provides 280+ hours of multi-camera 3D perception data for warehouses and smart spaces.
Before you read this
Two ideas this paper builds onRead it as
Wang and colleagues present Physical AI Smart Spaces, a benchmark for tracking people, robots and industrial vehicles in 3D across networks of fixed cameras.
The authors assembled 139 multi-camera scenes comprising 1,799 synchronized 1080p camera streams from warehouse-dominated indoor settings, alongside hospitals, retail venues and related spaces.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.