Event

Benchmark Evaluation of Nine Depth Estimation Models Released on Jetson AGX Orin

1 reports · 1 sources

Leaf Jiang benchmarked nine depth estimation model families on Jetson AGX Orin, covering monocular methods Depth Anything V2 and 3, Depth Pro, MoGe-2, Metric3D v2, and UniDepth V2, as well as stereo methods libSGM, NVIDIA ESS, and Fast-FoundationStereo, evaluating their accuracy, speed, and energy consumption. All methods run on Orin as TensorRT FP16 engines or native CUDA code, testing up to three input sizes (up to 5.7 MP), and metric accuracy is measured against Middlebury 2014 ground truth (non-occluded pixels with disparity below 256 px across 10 scenes). The author provides recommendations by frame-rate budget and publicly releases all code, per-scene results, and engine build scripts.