- depth-anything/Depth-Anything-V2-Small-hf
- INT8
- Orin NX 16GB
- Labels
- Image
- Depth
- Build
- Embedl-optimized
- Precision
- INT8
- Runtime
- TensorRT 10.16.2.10 with Embedl optimizations
- Input size
- 518×518
- Output
- Relative inverse depth
Depth Anything V2 Small turns a photo into relative inverse depth (larger values are nearer surfaces, not metres) and an illustrative point cloud.
Each example links its depth PNG from Embedl’s public media bucket; the point cloud stays in examples/*-output.json.
The upstream checkpoint is Apache-2.0; Embedl’s INT8 modifications are under the Embedl Models Community License v1.0. Example photos and depth images are CC BY 2.0, credited in NOTICE.
Methodology
- Latency: CUDA events, batch one, transfers excluded; 1s warmup, 5s measurement. Socket wall time includes image decoding, PNG/point cloud and JSON.
- Accuracy: Hypersim fixed 32-image test subset, eight scenes; not full benchmark. per-image least-squares scale+shift on inverse depth; 0.1<Z<10m, full image. Same square preprocessing for FP32 reference and TensorRT; no calibration/test scene overlap.
- Memory: Peak RSS of a fresh warmed serving subprocess; shared GPU RAM may not all appear in RSS. Engine activation memory reported separately in metrics.json.
- Size: the serialized TensorRT engine on disk.
- Baseline TensorRT: the FP16 model on upstream TensorRT 10.16.2.10, same preprocessing, same device.
Files
Variants
Every build of Depth Anything V2 Small on Orin NX 16GB.