Which Depth Estimator for Jetson AGX Orin? A 2026 Benchmark of Monocular and Stereo Methods

Leaf Jiang

cs.ROOct 19, 2026

Which Depth Estimator for Jetson AGX Orin? A 2026 Benchmark of Monocular and Stereo Methods

Authors: Leaf Jiang

Organizations: NODAR Inc.

Abstract

The Jetson AGX Orin is a common computer for robots and vehicles, yet measured accuracy, speed and energy of current depth estimators on it are hard to find. We benchmark nine model families as of October 2026: the monocular Depth Anything V2 and 3, Depth Pro, MoGe-2, Metric3D v2 and UniDepth V2, and the stereo methods libSGM, NVIDIA ESS and Fast-FoundationStereo. Every method runs on the Orin as a TensorRT FP16 engine or native CUDA code at up to three input sizes, up to the full 5.7 MP, and is scored for metric accuracy against Middlebury 2014 ground truth (non-occluded pixels of the 10 scenes whose disparity stays below 256 px), GPU time, energy and licence. Stereo is far more accurate. At 30 fps, NVIDIA's ESS reaches 0.70% AbsRel in 8.6 ms and libSGM 0.73% in 25 ms; with more compute, Fast-FoundationStereo reaches 0.33%. The best monocular models reach 12% with the true focal length (Depth Anything 3 Metric-L in 33 ms, Depth Pro in 803 ms), and 3.8-4.0% only with an oracle per-image scale (MoGe-2-L, UniDepth V2-B, Metric3D v2-L). Stereo error falls as the input resolution grows; no monocular model improves, and the metric scale of most drifts away from the truth outside their training resolution. At depth discontinuities, which AbsRel barely registers, the monocular models scatter 17-46% of the edge points between foreground and background, learned stereo 2-9% and libSGM at full resolution 1%. We give recommendations per frame-rate budget and release all code, per-scene results and engine build scripts.

Figures & tables

Method	Processing resolution	Full res.?	Max. disparity search range	Limiting factor	Licence (as stated by the authors)
Stereo
libSGM (Fixstars, 4 paths)	full ( $\approx$ 1990 $\times$ 2880); also half ( $\approx$ 995 $\times$ 1440) and quarter ( $\approx$ 497 $\times$ 720)	yes	256 levels at the processing resolution (256, $\approx$ 512 and $\approx$ 1024 px at full resolution)	—	Apache-2.0
ESS, Light ESS (NVIDIA Isaac ROS)	576 $\times$ 960 (ESS), 288 $\times$ 480 (Light ESS)	no	no fixed limit (regression)	fixed network input	NVIDIA TAO model EULA (commercial use permitted)
Fast-FoundationStereo (both checkpoints)	768 $\times$ 1152; also 512 $\times$ 768 and 384 $\times$ 576	no	$0.27\times$ input width: 320, 224 and 160 px ( $\approx$ 800 px at full resolution)	does not fit the 64 GB Orin at larger sizes: no TensorRT implementation for the cost-volume pooling at 960 $\times$ 1440; ONNX export runs out of host memory at 1280 $\times$ 1920	NVIDIA Source Code License (non-commercial)
C-Fast-FoundationStereo (commercial checkpoint)	768 $\times$ 1152; also 512 $\times$ 768 and 384 $\times$ 576	no	as Fast-FoundationStereo	as Fast-FoundationStereo	NVIDIA Open Model Agreement
Monocular

Table 1: Configuration of each method. All methods receive the same full-resolution (

{\approx}

5.7 MP) images and are scored against full-resolution ground truth; methods that cannot process 5.7 MP on the 64 GB Jetson AGX Orin run at their maximum input size. Full-resolution equivalents assume a 2880 px wide image.

Budget	Best stereo	Best monocular	Best monocular, oracle scale
	AbsRel (GPU time)	AbsRel (GPU time)	AbsRel (GPU time)
30 fps	ESS (NVIDIA), 576 $\times$ 960: 0.70% (9 ms)	Depth Anything 3 Metric-L, 350 $\times$ 504: 12.34% (33 ms)	Depth Anything 3 Metric-L, 350 $\times$ 504: 5.97% (33 ms)
10 fps	Fast-FoundationStereo 20-30-48, 384 $\times$ 576: 0.64% (68 ms)	Depth Anything 3 Metric-L, 350 $\times$ 504: 12.34% (33 ms)	UniDepth V2-B, 644 $\times$ 938: 3.90% (85 ms)
1 fps	C-Fast-FoundationStereo, 768 $\times$ 1152: 0.33% (760 ms)	Depth Pro, 1536 $\times$ 1536: 11.98% (803 ms)	MoGe-2-L, 700 $\times$ 1008: 3.78% (212 ms)
no limit	C-Fast-FoundationStereo, 768 $\times$ 1152: 0.33% (760 ms)	Depth Pro, 1536 $\times$ 1536: 11.98% (803 ms)	MoGe-2-L, 700 $\times$ 1008: 3.78% (212 ms)
lowest GPU, open source	libSGM, $\approx$ 497 $\times$ 720: 1.18% (6 ms)	Depth Anything V2-S (native), 518 $\times$ 770: 29.88% (12 ms)	Depth Anything V2-S (native), 518 $\times$ 770: 8.93% (12 ms)

Table 2: Recommendations: the most accurate configuration whose GPU time on one Jetson AGX Orin (MAXN, FP16) fits the frame-rate budget, by AbsRel on the non-occluded pixels of the 10 Middlebury 2014 scenes. Monocular models with the true focal length or intrinsics; the last column assumes the metric scale is supplied by an oracle (an upper bound). Last row: the configuration with the lowest GPU time among the methods whose licences permit commercial use and modification (libSGM, Depth Anything V2-S and 3, MoGe-2, Metric3D v2). Licences in table 1 .

Figure 1: Resolution scaling. AbsRel (non-occluded pixels, mean over the 10 scenes, log scale) against the number of pixels each method processes, for every method run at more than one input size. (a) As deployed. (b) Monocular models after per-image oracle median scaling (hollow markers), stereo unchanged. Stereo (solid lines): libSGM at quarter, half and full resolution and Fast-FoundationStereo (FFS) at

384\times 576

512\times 768

and

768\times 1152

, its largest input on the Orin (ESS has a fixed input). Monocular (dashed lines): each model from its native input to full resolution, with the true focal length or intrinsics. Stereo error falls with resolution; monocular error does not, and with an oracle scale it stays flat or rises.

Model	Input	MP	Scale GT/pred.	AbsRel [%]	Oracle AbsRel [%]	Density [%]
Depth Anything V2-S	518 $\times$ 770	0.40	1.45	29.9	8.9	100
Depth Anything V2-S	1008 $\times$ 1512	1.52	1.88	43.8	9.2	100
Depth Anything V2-S	1988 $\times$ 2884	5.73	2.44	56.0	9.7	100
Depth Anything V2-B	518 $\times$ 770	0.40	1.50	30.6	8.9	100
Depth Anything V2-B	1988 $\times$ 2884	5.73	1.83	41.2	10.6	100
Depth Anything V2-L	518 $\times$ 770	0.40	1.54	32.5	8.3	100

Table 3: Monocular models across input sizes. Scale: median over the scenes of the per-image median ratio of ground-truth to predicted depth (1 = unbiased,

>

1 = predicted too close). Density below 100% means that part of the scene is placed beyond the zero-disparity distance.

Figure 2: Point clouds at object edges, seen from above: depth $Z$ against lateral position $X$ of the points in a band of 40 image rows across the back slats of the Adirondack chair (top) and the front wheel of the motorcycle (bottom); the left column shows the rectified left image with the band in red and, below it, a magnified view of the band with some rows of context. Grey: ground truth. Each method is shown at its own output resolution (one point per output pixel); monocular depth with oracle scale. The monocular models connect foreground and background with ramps of “flying” points; Fast-FoundationStereo keeps dense, clean surfaces with a few flying points; libSGM at full resolution has the fewest flying points but noisier surfaces and missing points at the edges.

Method (output resolution)	Density at edges [%]	Flying points [%]	AbsRel at edges [%]	AbsRel elsewhere [%]
libSGM, full res.	45	1.1	4.33	0.59
libSGM, quarter res.	59	9.4	9.08	1.08
FFS 23-36-37, 768 $\times$ 1152	100	3.0	2.48	0.29
FFS 20-30-48, 768 $\times$ 1152	100	4.1	2.85	0.36
C-FFS, 768 $\times$ 1152	100	2.4	2.42	0.29
Depth Anything V2-S, 518 $\times$ 770	100	25.8	11.43	8.90

Table 4: Depth discontinuities, mean over the 10 scenes. Edge pixels: non-occluded ground-truth pixels whose

17\times 17

neighbourhood spans a depth ratio of at least 1.2 (1.7% of the scored pixels). Flying points: predicted edge pixels more than 5% in front of the farthest and behind the nearest ground-truth surface in that neighbourhood. Each method at its own output resolution (nearest-neighbour upsampling); monocular depth with oracle scale.

Figure 3: Depth maps (rows 1 and 3) and relative depth error on non-occluded pixels (rows 2 and 4) for Sword1 and Classroom1, with every method at full resolution or at its maximum or fixed input size. Stereo: raw outputs. FFS: Fast-FoundationStereo; DA: Depth Anything. Monocular: depth after per-image oracle median scaling, so the remaining error is structural, not a global scale error. Sword1 is the scene on which every method’s error is closest to its median over the 10 scenes; Classroom1 is added for its large planar surfaces. Numbers are the AbsRel of each panel. Depth uses one colour scale per scene (1st–99th percentile of the ground truth); values outside it are clipped. Grey: no ground truth, occluded, or no prediction (stereo holes are not filled). The first column of the error rows shows the scored (non-occluded) pixels in white.

Figure 4: As fig. 3 , at low resolution: libSGM at quarter resolution, Fast-FoundationStereo at $512\times 768$ and the monocular models at their native input, each as deployed, i.e. monocular depth with its own metric scale (last column: Depth Anything 3 with oracle scale for comparison). Column titles give the processing resolution and GPU time per frame on Jetson AGX Orin.

Figure 5: Depth error (AbsRel on non-occluded pixels, log scale) against GPU time per frame on Jetson AGX Orin, averaged over the 10 scenes. Filled markers are the methods as deployed (raw output; hole-filled results are in table 7 only); hollow markers are monocular methods with an oracle per-image scale. Upward triangles are Depth Anything V2 at full resolution, downward triangles at its native

518\times 770

input. The black stars are Depth Anything 3 Metric-L and the olive pentagons MoGe-2-S/B/L, the pink hexagons Metric3D v2-S/L and the blue circles UniDepth V2-S/B/L, with the ground-truth focal length or intrinsics; thin lines connect the native, twice-native and full-resolution runs of Depth Anything 3, MoGe-2-L, Metric3D v2-L and UniDepth V2-L (the smaller models at other sizes and MoGe-2 with its own focal estimate are in table 7 only). Green crosses are ESS and Light ESS, connected by a thin line; thin lines also connect the three input resolutions of libSGM (teal crosses) and of each Fast-FoundationStereo model. The small pink diamonds (C-Fast-FoundationStereo) coincide with the purple ones of the research checkpoint 23-36-37 . Labels give the processing resolution (height

\times

width). Vertical lines mark the per-frame budgets for 30 and 10 fps on one Orin.

Method	Resolution	Orins @10 fps	Orins @30 fps	$@30 fps,$ 20 per camera	$@30 fps,$ 500 per camera	Break-even camera @30 fps
Light ESS (NVIDIA)	288 $\times$ 480 (fixed)	0.03	0.08	293	1,253	–
libSGM	$\approx$ 497 $\times$ 720 (quarter)	0.06	0.18	592	1,552	–
ESS (NVIDIA)	576 $\times$ 960 (fixed)	0.09	0.26	816	1,776	–
libSGM	$\approx$ 995 $\times$ 1440 (half)	0.25	0.75	2,297	3,257	–
Fast-FoundationStereo 20-30-48	384 $\times$ 576	0.68	2.03	6,132	7,092	–
libSGM	full ( $\approx$ 1990 $\times$ 2880)	1.01	3.02	9,090	10,050	–

Table 5: System cost: cameras plus the fraction of Jetson AGX Orin 64GB GPU time used ($2,999 per module). Stereo uses two cameras and monocular one. “Orins” above 1 require multiple modules (or a larger GPU) for that frame rate. Break-even: the camera price below which adding the second camera and running Light ESS, the cheapest stereo configuration, is cheaper than the monocular model ( eq. 1 ). Monocular models with their own focal-length estimate have the same cost as with the true one and are omitted.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Scene	Size ( $W\times H$ )	ndisp	min	median	99 %	max	max [px]	[m]	[%]	[%]
			GT disparity [px]				Non-occ.	Depth	Occl.	Out
Adirondack	2880 $\times$ 1988	280	30	108	219	243	243	1.62–3.07	6.1	2.0
Backpack	2940 $\times$ 2016	260	32	130	214	224	224	2.22–3.36	8.4	3.1
Bicycle1	2988 $\times$ 2008	180	56	73	127	145	145	2.94–4.08	6.3	2.6
Classroom1	3000 $\times$ 1920	250	16	55	203	208	208	2.98–7.63	2.7	1.8
Motorcycle	2964 $\times$ 2000	270	28	152	231	240	240	2.11–5.04	7.9	3.1

Appendix

Table 6: The 10 Middlebury 2014 scenes used (left image thumbnails). Disparities in full-resolution pixels; depth from

Z=fB/(d+d_{\text{offs}})

Parameter	Value	Note
$P_{1}$	4 (full), 6 (half), 3 (quarter)	tuned; default 10
$P_{2}$	224 at every resolution	tuned; default 120; 224 is the 8-bit maximum
Aggregation paths	4 (horizontal and vertical)	default 8
Sub-pixel output	on (4 fractional bits)	default off
Disparity levels	256	maximum supported
Minimum disparity	0	default

Appendix

Table 8: libSGM (version 3.1.0, commit e4c669b ) parameters.

Metric Depth Estimation

Metric depth estimation predicts scene distances in physical units from images, supporting robotics, mapping, augmented reality, and 3D reconstruction.

See all 64 papers

AsyncMDE: Real-Time Monocular Depth Estimation via Asynchronous Spatial Memory
Mar 11, 2026Lianjie Ma, Yuquan Li, Bingzheng Jiang +3Edge Inference Monocular Depth Estimation
School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China. · Yichang Testing Technology Research Institute, Yichang, China. · School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan, China.
DepthAgent: Towards Better Universal Depth Estimation via Sample-wise Expert Selection
May 22, 2026Jie Zhu, Girish Chandar Ganesan, Xiaoming LiuMetric Depth Estimation Monocular Depth Estimation
Michigan State University · University of North Carolina at Chapel Hill
Depth-Aware Rover: A Study of Edge AI and Monocular Vision for Real-World Implementation
Apr 24, 2026Lomash Relia, Jai G Singla, Amitabh +1Metric Depth Estimation Robot Navigation
Department of Computer Engineering Devang Patel Institute of Advanced Technology and Research Anand, India · Space Applications Centre Indian Space Research Organization Ahmedabad, India

Mar 11, 2026J/K move · Enter open · S save

AsyncMDE: Real-Time Monocular Depth Estimation via Asynchronous Spatial Memory

Lianjie Ma, Yuquan Li, Bingzheng Jiang, Ziming Zhong+2

School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China. · Yichang Testing Technology Research Institute, Yichang, China. · School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan, China.

Edge Inference Monocular Depth Estimation

Open paper AskPDF arXiv

Foundation-model-based monocular depth estimation offers a viable alternative to active sensors for robot perception, yet its computational cost often prohibits deployment on edge platforms. Existing methods perform independent per-frame inference, wasting the substantial computational redundancy between adjacent viewpoints in continuous robot operation. This paper presents AsyncMDE, an asynchronous depth perception system consisting of a frozen foundation model and a lightweight fast path that amortizes the foundation model's computational cost over time. The foundation model periodically produces high-quality spatial features in the background, while the lightweight fast path runs asynchronously in the foreground, fusing cached memory with current observations through complementary fusion, outputting depth estimates, and autoregressively updating memory. This enables cross-frame feature reuse with bounded accuracy degradation. With 3.83M trainable fast-path parameters and a 97.5M frozen slow path, AsyncMDE's fast path operates at 237 FPS on an RTX 4090, recovering 77% of the accuracy gap to the foundation model. Across indoor static, dynamic, and synthetic extreme-motion benchmarks, AsyncMDE degrades predictably and reaches 161 FPS fast-path inference on a TensorRT-optimized Jetson AGX Orin, supporting real-time edge deployment.

Abstract

Metric Depth Estimation

Explore similar work

AsyncMDE: Real-Time Monocular Depth Estimation via Asynchronous Spatial Memory

DepthAgent: Towards Better Universal Depth Estimation via Sample-wise Expert Selection

Depth-Aware Rover: A Study of Edge AI and Monocular Vision for Real-World Implementation