Embedl's build nvidia/sparse4d of the original model nvidia/sparse4d_rn50
Embedl-optimized
- nvidia/tao/sparse4d_rn50:trainable_v2.2
- FP8
- AGX Thor
- Labels
- Image
- Detection
- Build
- Embedl-optimized
- Precision
- FP8 + FP16
- Runtime
- TensorRT 10.13.3.9 with Embedl optimizations
- Cameras
- 13 views, 540×960 each
- Output
- 3D boxes with track IDs
Sparse4D (NVIDIA TAO, ResNet-50) detects and tracks people and vehicles in 3D across 13 warehouse cameras at once.
Methodology
- Latency: one frame of all 13 cameras through the TensorRT 10.13 engine in one pass (static shape), batch 1, median; the temporal runtime between frames is not included.
- Accuracy: all 9,000 frames of PhysicalAI-SmartSpaces MTMC 2025 val/Warehouse_015 (77 tracks: Person, Forklift, Transporter, NovaCarter) through the same runtime, scored with NVIDIA’s spatialai-data-utils HOTA evaluator (3D-IoU matching). No camera grouping, so not comparable with NVIDIA’s published grouped-protocol HOTA. Calibration used another scene (Warehouse_016).
- Size: the serialized TensorRT engine on disk.
- Baseline TensorRT: the same model, unquantized, through the same torch.export → ONNX path, built FP16 with TensorRT on the same device. FP16 builds of NVIDIA’s TAO-exported ONNX return NaN on Thor (a TensorRT GridSample defect), so that path is the baseline.