- huggingface
- INT8
- L4
- Labels
- Image
- Pose
- Build
- Embedl-optimized
- Precision
- INT8 + FP16
- Runtime
- TensorRT 11.3.0.99 with Embedl optimizations
- People
- Up to 5 per photo
- Output
- A 3D body mesh per person
Example
Ran on NVIDIA L4 (sm_89), TensorRT 11.3.0.99, during the build
COCO val2017 000000229553 (CC BY 2.0 (COCO val2017 licence 4))
--image 
3D body
One photo in, a 3D body mesh for every person in it out: Meta's SAM 3D Body (DINOv3 ViT-H+/16 encoder), with a YOLOX-S person detector in front of it.
Derivative of facebook/sam-3d-body-dinov3 (SAM 3D Body License) and of the DINOv3 weights (DINOv3 License); Embedl's quantized encoder and the decoder export are under the Embedl Models Community Licence. YOLOX-S: Megvii, Apache-2.0. The MHR keypoint names and links: facebookresearch/sam-3d-body. NOTICE says the same beside the files.
Methodology
- Latency: on the NVIDIA L4, batch 1. Engines: GPU time of one execution between CUDA events, copies excluded. Photo to 3D body: wall time of serve.py's whole answer (detector, crops, engines with copies, packing), median of 10 after a warm-up.
- Accuracy: each photo through serve.py twice, on these engines and on FP32 in onnxruntime (CPU): the published decoder and detector graphs and Embedl's encoder graph with its Q/DQ stripped, unquantized; the reference's person boxes given to both. The mean distance of the 18,439 vertices per person, then over people; keypoints likewise. COCO val2017 photos (CC BY 2.0).
- Size: the detector, encoder and decoder TensorRT engines on disk.
- FP32 reference: Embedl’s encoder graph unquantized with the published decoder and detector graphs, in onnxruntime FP32 on the CPU.
- Baseline TensorRT: the FP16 build of the same model with TensorRT 11.3.0.99, same device, same measurement.
Files
Variants
Every build of SAM 3D Body on L4.