LuMon: A Comprehensive Benchmark and Development Suite with Novel Datasets for Lunar Monocular Depth Estimation

Aytaç Sekmen; Fatih Emre Gunes; Furkan Horoz; Hüseyin Umut Işık; Mehmet Alp Ozaydin; Onur Altay Topaloglu; Şahin Umutcan Üstündaş; Yurdasen Alp Yeni

cs.CVApr 10, 2026

LuMon: A Comprehensive Benchmark and Development Suite with Novel Datasets for Lunar Monocular Depth Estimation

Authors: Aytaç Sekmen, Fatih Emre Gunes, Furkan Horoz, Hüseyin Umut Işık, Mehmet Alp Ozaydin, Onur Altay Topaloglu, Şahin Umutcan Üstündaş, Yurdasen Alp Yeni,

Organizations: Department of Computer Engineering, Middle East Technical University, Ankara, Turkey · ROMER Middle East Technical University, Ankara, Turkey · Department of Aerospace Engineering, Middle East Technical University, Ankara, Turkey

Abstract

Monocular Depth Estimation (MDE) is crucial for autonomous lunar rover navigation using electro-optical cameras. However, deploying terrestrial MDE networks to the Moon brings a severe domain gap due to harsh shadows, textureless regolith, and zero atmospheric scattering. Existing evaluations rely on analogs that fail to replicate these conditions and lack actual metric ground truth. To address this, we present LuMon, a comprehensive benchmarking framework to evaluate MDE methods for lunar exploration. We introduce novel datasets featuring high-quality stereo ground truth depth from the real Chang'e-3 mission and the CHERI dark analog dataset. Utilizing this framework, we conduct a systematic zero-shot evaluation of state-of-the-art architectures across synthetic, analog, and real datasets. We rigorously assess performance against mission critical challenges like craters, rocks, extreme shading, and varying depth ranges. Furthermore, we establish a sim-to-real domain adaptation baseline by fine tuning a foundation model on synthetic data. While this adaptation yields drastic in-domain performance gains, it exhibits minimal generalization to authentic lunar imagery, highlighting a persistent cross-domain transfer gap. Our extensive analysis reveals the inherent limitations of current networks and sets a standard foundation to guide future advancements in extraterrestrial perception and domain adaptation.

Figures & tables

Figure 1: LuMon uses real and synthetic lunar surface datasets. We utilize leading metric and relative monocular depth estimation models to extract depthmaps, and evaluate their performance in zero-shot lunar settings. We measure the overall performance of the networks in terms of accuracy and efficiency, their accuracy in challenging conditions, and effectiveness in downstream tasks. Finally, we perform fine-tuning to compare it against zero-shot performance.

Dataset	Domain	Seg. Masks	GT Source	Baseline	Resolutions	Max Depth
LunarSim	Sim.	✗	Sim.	12 cm	$1280\times 720$	-
LuSNAR	Sim.	✓	Sim.	31 cm	$1024\times 1024$	50 m
Etna-S3LI	Analog	✗	LiDAR	20 cm	$688\times 512$	30 m
Etna-LRNT	Analog	✗	Stereo	9 cm	$1292\times 964$	15 m
CHERI	Analog	✗	Stereo	11.2 cm	$1280\times 720$	17 m
Chang’e-3	In-Situ	✗	Stereo	27 cm	$2352\times 1728$	25 m

Table 1: A summary of the datasets used. GT : Ground Truth; Seg. Masks : Segmentation Masks; Sim : Simulation.

Type	Model	LunarSim			LuSNAR			Etna-LRNT			Etna-S3LI			Chang’e-3			CHERI
Type	Model	$\delta_{1}\uparrow$	A.Rel $\downarrow$	RMSE $\downarrow$	$\delta_{1}\uparrow$	A.Rel $\downarrow$	RMSE $\downarrow$	$\delta_{1}\uparrow$	A.Rel $\downarrow$	RMSE $\downarrow$	$\delta_{1}\uparrow$	A.Rel $\downarrow$	RMSE $\downarrow$	$\delta_{1}\uparrow$	A.Rel $\downarrow$	RMSE $\downarrow$	$\delta_{1}\uparrow$	A.Rel $\downarrow$	RMSE $\downarrow$
Metric	DAv2-FT (Ours)	0.84	0.13	13.96	0.93	0.07	1.07	1.00	0.05	0.13	0.84	0.13	1.30	0.92	0.08	0.70	0.36	0.50	3.90
	DepthAnything v2 [ 49 ]	0.56	0.26	15.18	0.31	0.45	2.89	1.00	0.06	0.15	0.88	0.12	1.22	0.95	0.06	0.56	0.32	0.54	4.14
	DepthAnything 3 [ 11 ]	0.63	0.22	14.15	0.53	0.29	1.99	1.00	0.04	0.11	0.91	0.10	1.12	0.96	0.05	0.44	0.34	0.51	3.97
	MoGe-2 [ 44 ]	0.77	0.15	12.17	0.53	0.29	1.80	0.99	0.05	0.15	0.91	0.10	1.08	0.95	0.06	0.45	0.32	0.54	4.12
	Depth Pro [ 4 ]	0.50	0.29	17.57	0.33	0.45	2.70	0.73	241.71	450.05	0.89	0.11	1.11	0.89	0.10	0.89	0.32	0.55	4.18

Table 2: Exp. 1: Evaluation on LunarSim, LuSNAR, Etna, S3LI, Chang’e 3, and Cheri datasets. Red and blue respectively indicate best and second-best results.

Type	Model	FPS $\uparrow$	VRAM (GB) $\downarrow$	TFLOPs $\downarrow$
Metric	DAv2-FT (Ours)	4.64	1.926	1.309
	DepthAnything v2	4.54	1.926	1.309
	DepthAnything 3	24.65	2.007	0.658
	MoGe-2	2.63	2.234	2.150
	Depth Pro	0.91	7.949	9.677
	Metric-3D v2	2.47	2.990	2.870

Table 3: Efficiency metric results. Red indicates the best result and blue indicates the second best.

Type	Model	Regolith		Rocks		Craters
Type	Model	$\delta_{1}\uparrow$	A.Rel $\downarrow$	$\delta_{1}\uparrow$	A.Rel $\downarrow$	$\delta_{1}\uparrow$	A.Rel $\downarrow$
Metric	DAv2-FT (Ours)	0.93	0.07	0.92	0.09	0.87	0.10
	DepthAnything v2 [ 49 ]	0.30	0.46	0.44	0.28	0.23	0.37
	DepthAnything 3 [ 11 ]	0.52	0.29	0.65	0.19	0.37	0.27
	MoGe-2 [ 44 ]	0.52	0.30	0.69	0.18	0.40	0.27
	Depth Pro [ 4 ]	0.32	0.47	0.45	0.28	0.23	0.37

Table 4: Exp. 2: Results on semantically different regions on LuSNAR dataset (Regolith, Rock and Craters). Red indicates the best result and blue indicates the second best.

Type	Model	LunarSim			LuSNAR			Etna-LRNT			Etna-S3LI			Chang’e-3			CHERI
Type	Model	$\delta_{1}\uparrow$	A.Rel $\downarrow$	RMSE $\downarrow$	$\delta_{1}\uparrow$	A.Rel $\downarrow$	RMSE $\downarrow$	$\delta_{1}\uparrow$	A.Rel $\downarrow$	RMSE $\downarrow$	$\delta_{1}\uparrow$	A.Rel $\downarrow$	RMSE $\downarrow$	$\delta_{1}\uparrow$	A.Rel $\downarrow$	RMSE $\downarrow$	$\delta_{1}\uparrow$	A.Rel $\downarrow$	RMSE $\downarrow$
Metric	DAv2-FT (Ours)	0.79	0.15	15.10	0.84	0.11	0.95	1.00	0.06	0.10	0.46	0.29	2.88	0.89	0.10	0.73	0.16	0.51	4.76
	DepthAnything v2 [ 49 ]	0.47	0.29	16.39	0.20	0.66	2.86	0.99	0.07	0.12	0.51	0.33	2.85	0.94	0.07	0.59	0.15	0.72	5.18
	DepthAnything 3 [ 11 ]	0.54	0.26	15.77	0.34	0.40	2.15	1.00	0.05	0.07	0.53	0.46	3.58	0.94	0.06	0.45	0.18	0.62	4.75
	MoGe-2 [ 44 ]	0.68	0.18	14.42	0.35	0.41	1.90	0.99	0.05	0.08	0.56	0.38	3.22	0.93	0.07	0.46	0.15	0.73	5.20
	Depth Pro [ 4 ]	0.38	0.35	21.33	0.23	0.68	2.83	0.39	559.37	818.97	0.59	0.25	2.36	0.85	0.11	0.87	0.13	0.78	5.27

Table 5: Exp. 2: Performances on shaded regions. Red indicates the best result and blue indicates the second best.

		Near					Far
Type	Model	LuSNAR	Etna-LRNT	Etna-S3LI	Chang’e-3	CHERI	LuSNAR	Etna-LRNT	Etna-S3LI	Chang’e-3	CHERI
Metric	Dav2-FT (Ours)	0.07	0.05	0.18	0.34	1.55	0.09	0.20	0.29	0.15	0.23
	DepthAnything v2 [ 49 ]	0.49	0.06	0.15	0.18	1.74	0.20	0.30	0.29	0.16	0.25
	DepthAnything 3 [ 11 ]	0.31	0.04	0.14	0.20	1.58	0.16	0.29	0.27	0.14	0.24
	MoGe-2 [ 44 ]	0.32	0.05	0.13	0.22	1.72	0.13	0.56	0.26	0.14	0.25
	Depth Pro [ 4 ]	0.50	241.77	0.13	0.34	1.77	0.18	0.69	0.27	0.20	0.25

Table 6: Exp. 3: Evaluation of Absolute Relative Error (A.Rel

\downarrow

) on all datasets except LunarSim across Near and Far distances. Red : Best, blue : Second-best.

Type	Model	Scale	Real%	MetSup	Dominant Scene
Metric	DAv2-FT (Ours)	62.51M	99%	1%*	Gen. Open-World, LuSNAR.
	DepthAnything v2 [ 49 ]	62.5M	99%	1%*	General Open-World
	DepthAnything 3 [ 11 ]	69M	97%	100%*	General Open-World
	MoGe-2 [ 44 ]	8.8M	74%	98%	Indoor
	Depth Pro [ 4 ]	3M	15%	98%	Synthetic Simulation
	Metric-3D v2 [ 19 ]	16.2M	99%	85%	Indoor, Outdoor Driving

Table 7: Exp 4: Comparison of Training Data Scales and Composition. Scale: Total data size. Real %: Percentage of real data. MetSup: Percentage of data with metric supervision.

Panel of Figure 2 — Figure 2: Effect of training dataset scale for one dataset. (a) Scale vs. RMSE. (b) Spearman correlation of scale vs. $\delta_{1}$ ( $*$ : $p<0.05$ , $**$ : $p<0.01$ ).

Figure 3: An example scene from Chang’e 3 ( RGB GT ), its ground-truth depth map ( GT Depth ), and the DepthAnything v2 predictions (base and fine-tuned). Note: Differences in visual boundaries between the RGB GT and depth maps result from the evaluation procedure (Section 3.5 ); where ground-truth maps and predictions are masked by maximum depth boundaries and invalid pixels. Additionally, the depth maps generated by the other models can be found in the supplementary material.

Type	Method	Med. Err. $\downarrow$		Pose Err. AUC (%) $\uparrow$
Type	Method	$\epsilon_{R}\ (^{\circ})$	$\epsilon_{t}\ (^{\circ})$	$@5^{\circ}$	$@10^{\circ}$	$@20^{\circ}$
Metric	DAv2-FT (Ours)	0.12	1.59	57.70	78.05	89.03
	DepthAnything v2 [ 49 ]	0.15	1.89	56.22	77.27	88.64
	DepthAnything 3 [ 11 ]	0.14	2.17	53.28	76.04	88.02
	MoGe-2 [ 44 ]	0.15	1.68	58.40	78.52	89.26
	Depth Pro [ 4 ]	0.22	2.84	41.12	69.78	84.89

Table 8: Exp. 6: Relative Pose Estimation using MADPose on LuSNAR test-set. Red : Best, blue : Second-best.

Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.84	0.97	0.99	0.13	2.04	13.96	7.04	15.95	2.56	130.43
	DepthAnything v2 [ 49 ]	0.56	0.81	0.89	0.26	3.62	15.18	10.04	95.84	3.57	106.37
	DepthAnything 3 [ 11 ]	0.63	0.85	0.91	0.22	2.84	14.15	8.89	55.14	3.19	157.79
	MoGe-2 [ 44 ]	0.77	0.93	0.97	0.15	1.76	12.17	6.80	24.89	2.02	140.01
	Depth Pro [ 4 ]	0.50	0.79	0.88	0.29	4.64	17.57	11.45	84.24	3.58	152.95
	Metric-3D v2 [ 19 ]	0.75	0.91	0.97	0.16	1.99	12.53	7.10	29.10	2.47	72.61

Appendix

Table S1 : Exp. 1: Results on LunarSim dataset. Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.93	0.99	1.00	0.07	0.07	1.07	0.41	9.13	1.09	2.53
	DepthAnything v2 [ 49 ]	0.31	0.63	0.73	0.45	0.92	2.89	1.77	342.41	7.10	86.42
	DepthAnything 3 [ 11 ]	0.53	0.76	0.85	0.29	0.44	1.99	1.13	121.12	2.26	38.61
	MoGe-2 [ 44 ]	0.53	0.75	0.84	0.29	0.40	1.80	1.07	136.76	2.11	36.22
	Depth Pro [ 4 ]	0.33	0.64	0.72	0.45	0.91	2.70	1.69	374.12	6.65	99.65
	Metric-3D v2 [ 19 ]	0.61	0.79	0.86	0.25	0.33	1.71	0.97	108.62	1.92	45.65

Appendix

Table S2 : Exp. 1: Results on LuSNAR dataset. Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	1.00	1.00	1.00	0.05	0.01	0.13	0.10	5.56	1.90	41.06
	DepthAnything v2 [ 49 ]	1.00	1.00	1.00	0.06	0.01	0.15	0.12	7.01	3.81	4.84
	DepthAnything 3 [ 11 ]	1.00	1.00	1.00	0.04	0.01	0.11	0.08	4.91	2.39	36.57
	MoGe-2 [ 44 ]	0.99	1.00	1.00	0.05	0.01	0.16	0.11	6.37	3.33	13.87
	Depth Pro [ 4 ]	0.73	0.93	0.95	241.71	2415143.42	450.05	449.98	18.12	3.01	70.35
	Metric-3D v2 [ 19 ]	1.00	1.00	1.00	0.04	0.00	0.10	0.07	4.40	2.44	6.42

Appendix

Table S3 : Exp. 1: Results on Etna-LRNT dataset. Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$
Metric	DAv2-FT	0.84	0.97	0.99	0.13	0.26	1.30	0.94	15.97
	DepthAnything v2 [ 49 ]	0.88	0.98	0.99	0.12	0.21	1.22	0.85	34.64
	DepthAnything 3 [ 11 ]	0.91	0.98	0.99	0.10	0.18	1.12	0.74	38.88
	MoGe-2 [ 44 ]	0.91	0.99	1.00	0.10	0.17	1.08	0.73	28.97
	Depth Pro [ 4 ]	0.89	0.98	1.00	0.11	0.18	1.11	0.78	16.58
	Metric-3D v2 [ 19 ]	0.87	0.98	0.99	0.12	0.22	1.19	0.83	27.04

Appendix

Table S4 : Exp. 1: Results on Etna-S3LI dataset. Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.92	1.00	1.00	0.08	0.08	0.70	0.51	10.45	3.34	401.86
	DepthAnything v2 [ 49 ]	0.95	1.00	1.00	0.06	0.05	0.56	0.41	8.30	3.42	71.96
	DepthAnything 3 [ 11 ]	0.96	1.00	1.00	0.05	0.03	0.44	0.30	7.77	3.00	242.66
	MoGe-2 [ 44 ]	0.95	1.00	1.00	0.06	0.04	0.45	0.32	7.40	2.76	126.03
	Depth Pro [ 4 ]	0.89	0.98	1.00	0.10	0.17	0.89	0.67	13.51	2.85	204.48
	Metric-3D v2 [ 19 ]	0.95	1.00	1.00	0.05	0.04	0.43	0.30	7.12	3.27	30.09

Appendix

Table S5 : Exp. 1: Results on Chang’e-3 dataset. Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.36	0.70	0.84	0.50	2.55	3.90	3.31	50.74	2.46	84.84
	DepthAnything v2 [ 49 ]	0.32	0.65	0.83	0.54	2.88	4.14	3.56	50.62	3.23	94.51
	DepthAnything 3 [ 11 ]	0.34	0.69	0.85	0.51	2.63	3.97	3.39	48.38	2.65	69.40
	MoGe-2 [ 44 ]	0.32	0.65	0.83	0.54	2.84	4.12	3.54	50.92	3.24	174.46
	Depth Pro [ 4 ]	0.32	0.64	0.82	0.55	2.88	4.18	3.60	51.01	3.52	41.94
	Metric-3D v2 [ 19 ]	0.33	0.67	0.84	0.53	2.83	4.09	3.51	51.84	2.99	130.94

Appendix

Table S6 : Exp. 1: Results on CHERI dataset. Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.79	0.97	0.99	0.15	3.23	15.10	10.58	12.90	2.70	71.79
	DepthAnything v2 [ 49 ]	0.47	0.72	0.84	0.29	4.33	16.39	12.44	75.94	4.03	96.96
	DepthAnything 3 [ 11 ]	0.54	0.75	0.86	0.26	3.67	15.77	11.29	50.73	3.52	125.68
	MoGe-2 [ 44 ]	0.68	0.90	0.96	0.17	2.79	14.42	10.28	20.76	2.71	76.19
	Depth Pro [ 4 ]	0.38	0.62	0.75	0.35	6.84	21.33	15.72	91.85	4.09	116.75
	Metric-3D v2 [ 19 ]	0.62	0.85	0.94	0.20	3.11	14.35	9.59	25.92	3.32	41.03

Appendix

Table S7 : Exp. 2: Results on LunarSim dataset (shaded regions). Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.84	0.96	1.00	0.11	0.10	0.95	0.49	7.94	1.57	27.25
	DepthAnything v2 [ 49 ]	0.20	0.37	0.45	0.66	1.39	2.86	2.11	407.91	7.59	62.95
	DepthAnything 3 [ 11 ]	0.34	0.54	0.68	0.40	0.64	2.15	1.39	131.02	3.50	42.53
	MoGe-2 [ 44 ]	0.35	0.53	0.65	0.41	0.59	1.90	1.31	147.67	3.34	44.50
	Depth Pro [ 4 ]	0.23	0.39	0.46	0.68	1.40	2.83	2.04	430.27	7.58	67.47
	Metric-3D v2 [ 19 ]	0.39	0.57	0.69	0.38	0.57	1.92	1.30	120.21	3.56	52.68

Appendix

Table S8 : Exp. 2: Results on LuSNAR dataset (shaded regions). Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	1.00	1.00	1.00	0.06	0.01	0.10	0.09	1.25	2.85	37.38
	DepthAnything v2 [ 49 ]	0.99	1.00	1.00	0.07	0.01	0.12	0.12	2.24	2.60	12.35
	DepthAnything 3 [ 11 ]	1.00	1.00	1.00	0.05	0.00	0.07	0.07	1.50	2.55	27.67
	MoGe-2 [ 44 ]	0.99	1.00	1.00	0.05	0.01	0.08	0.08	1.61	2.81	21.85
	Depth Pro [ 4 ]	0.39	0.83	0.92	559.37	5590194.98	818.97	818.96	4.29	4.19	34.97
	Metric-3D v2 [ 19 ]	1.00	1.00	1.00	0.05	0.00	0.07	0.07	1.51	2.69	23.61

Appendix

Table S9 : Exp. 2: Results on Etna-LRNT dataset (shaded regions). Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$
Metric	DAv2-FT	0.46	0.70	0.84	0.29	1.14	2.88	2.59	15.79
	DepthAnything v2 [ 49 ]	0.47	0.63	0.68	0.43	2.36	3.54	3.27	109.25
	DepthAnything 3 [ 11 ]	0.53	0.65	0.68	0.46	2.90	3.58	3.30	80.69
	MoGe-2 [ 44 ]	0.56	0.69	0.73	0.38	2.30	3.22	2.92	95.08
	Depth Pro [ 4 ]	0.59	0.79	0.87	0.25	0.99	2.36	2.11	25.29
	Metric-3D v2 [ 19 ]	0.68	0.87	0.93	0.23	1.01	2.35	1.95	91.02

Appendix

Table S10 : Exp. 2: Results on Etna-S3LI dataset (shaded regions). Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.89	0.99	1.00	0.10	0.10	0.73	0.55	10.59	3.74	241.28
	DepthAnything v2 [ 49 ]	0.94	1.00	1.00	0.07	0.06	0.59	0.44	8.36	3.43	50.51
	DepthAnything 3 [ 11 ]	0.94	1.00	1.00	0.06	0.04	0.45	0.32	9.67	3.09	119.34
	MoGe-2 [ 44 ]	0.93	1.00	1.00	0.07	0.05	0.46	0.34	7.81	2.71	95.83
	Depth Pro [ 4 ]	0.85	0.98	1.00	0.11	0.18	0.87	0.69	15.04	3.08	139.39
	Metric-3D v2 [ 19 ]	0.94	1.00	1.00	0.06	0.04	0.44	0.32	7.41	2.71	44.76

Appendix

Table S11 : Exp. 2: Results on Chang’e-3 dataset (shaded regions). Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.16	0.50	0.82	0.51	2.51	4.76	4.27	58.39	3.28	10.11
	DepthAnything v2 [ 49 ]	0.15	0.39	0.76	0.72	4.12	5.18	4.81	44.18	3.69	11.32
	DepthAnything 3 [ 11 ]	0.18	0.49	0.79	0.62	3.17	4.75	4.37	39.39	3.06	9.37
	MoGe-2 [ 44 ]	0.15	0.40	0.76	0.73	4.26	5.20	4.83	44.93	3.01	9.48
	Depth Pro [ 4 ]	0.13	0.45	0.73	0.78	4.73	5.27	4.93	47.67	3.46	9.83
	Metric-3D v2 [ 19 ]	0.16	0.43	0.78	0.68	3.83	5.06	4.67	46.47	2.88	11.23

Appendix

Table S12 : Exp. 2: Results on CHERI dataset (shaded regions). Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.93	0.99	1.00	0.07	0.04	0.75	0.30	8.59	0.34	2.92
	DepthAnything v2 [ 49 ]	0.30	0.61	0.71	0.46	0.84	2.24	1.52	350.76	7.34	104.45
	DepthAnything 3 [ 11 ]	0.52	0.75	0.84	0.29	0.40	1.63	0.99	123.58	1.98	44.95
	MoGe-2 [ 44 ]	0.52	0.74	0.83	0.30	0.38	1.53	0.97	139.21	2.51	39.42
	Depth Pro [ 4 ]	0.32	0.63	0.71	0.47	0.85	2.25	1.50	382.05	7.78	99.93
	Metric-3D v2 [ 19 ]	0.60	0.78	0.85	0.26	0.30	1.34	0.84	111.11	1.87	56.49

Appendix

Table S13 : Exp. 2: Results on Regolith regions on LuSNAR dataset. Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.92	0.99	1.00	0.09	0.50	3.56	2.56	10.36	2.87	12.83
	DepthAnything v2 [ 49 ]	0.44	0.83	0.97	0.28	2.21	7.15	6.13	18.04	3.38	11.37
	DepthAnything 3 [ 11 ]	0.65	0.95	0.99	0.19	1.28	5.76	4.60	15.00	3.33	17.25
	MoGe-2 [ 44 ]	0.69	0.96	1.00	0.18	1.14	5.36	4.30	13.06	2.79	7.57
	Depth Pro [ 4 ]	0.45	0.83	0.96	0.28	2.46	7.49	6.41	18.03	2.66	6.00
	Metric-3D v2 [ 19 ]	0.80	0.97	0.99	0.15	1.00	4.76	3.70	12.87	3.16	9.12

Appendix

Table S14 : Exp. 2: Results on Rocks regions on LuSNAR dataset. Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.87	0.98	1.00	0.10	0.77	4.52	3.83	8.48	2.05	10.06
	DepthAnything v2 [ 49 ]	0.23	0.45	0.73	0.37	6.69	15.30	14.44	15.69	2.67	6.68
	DepthAnything 3 [ 11 ]	0.37	0.75	0.90	0.27	4.16	11.53	10.82	12.38	2.25	6.50
	MoGe-2 [ 44 ]	0.40	0.71	0.88	0.27	4.45	11.79	11.05	12.21	2.14	11.12
	Depth Pro [ 4 ]	0.23	0.44	0.72	0.37	7.14	15.72	14.82	15.87	2.29	5.56
	Metric-3D v2 [ 19 ]	0.41	0.75	0.90	0.27	3.87	11.11	10.31	19.84	2.46	9.85

Appendix

Table S15 : Exp. 2: Results on Craters regions on LuSNAR dataset. Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.92	0.98	1.00	0.07	0.02	0.16	0.14	6.96	0.41	7.46
	DepthAnything v2 [ 49 ]	0.30	0.55	0.64	0.49	0.68	1.13	0.99	379.83	7.18	158.41
	DepthAnything 3 [ 11 ]	0.48	0.69	0.80	0.31	0.31	0.70	0.61	130.89	2.85	84.58
	MoGe-2 [ 44 ]	0.48	0.68	0.78	0.32	0.31	0.70	0.62	148.40	3.24	102.99
	Depth Pro [ 4 ]	0.33	0.55	0.63	0.50	0.73	1.15	0.99	413.67	7.55	165.65
	Metric-3D v2 [ 19 ]	0.55	0.73	0.81	0.28	0.26	0.64	0.56	119.17	3.30	91.01

Appendix

Table S16 : Exp. 3: Results on Near (0–5m) distances on LuSNAR dataset. Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.99	1.00	1.00	0.07	0.10	0.88	0.60	4.47	1.17	36.08
	DepthAnything v2 [ 49 ]	0.20	0.83	0.99	0.40	1.49	3.29	3.13	7.61	2.03	9.27
	DepthAnything 3 [ 11 ]	0.58	0.96	1.00	0.25	0.68	2.23	2.00	6.96	1.21	33.29
	MoGe-2 [ 44 ]	0.59	0.96	1.00	0.25	0.64	2.12	1.97	5.86	0.92	18.43
	Depth Pro [ 4 ]	0.19	0.87	0.99	0.39	1.44	3.21	3.05	7.35	1.31	26.11
	Metric-3D v2 [ 19 ]	0.74	0.98	1.00	0.19	0.45	1.77	1.53	8.70	1.34	13.91

Appendix

Table S17 : Exp. 3: Results on Medium (5–15m) distances on LuSNAR dataset. Red indicates the best result and blue indicates the second best.

Type	Method	$\delta_{1}$ $\uparrow$	$\delta_{2}$ $\uparrow$	$\delta_{3}$ $\uparrow$	A.Rel $\downarrow$	Sq.Rel $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	SILog $\downarrow$	E.Acc $\downarrow$	E.Comp $\downarrow$
Metric	DAv2-FT	0.93	0.99	0.99	0.09	0.49	3.70	2.60	11.91	1.52	11.10
	DepthAnything v2 [ 49 ]	0.60	0.85	0.95	0.20	2.28	8.78	6.50	25.11	2.77	14.19
	DepthAnything 3 [ 11 ]	0.75	0.93	0.98	0.16	1.31	6.27	4.59	19.20	1.96	41.20
	MoGe-2 [ 44 ]	0.82	0.96	0.99	0.13	0.97	5.57	3.91	16.50	1.13	20.87
	Depth Pro [ 4 ]	0.65	0.88	0.97	0.18	1.91	8.13	5.86	22.27	1.42	28.95
	Metric-3D v2 [ 19 ]	0.81	0.96	0.99	0.14	1.02	5.46	4.06	18.08	1.77	11.11

Appendix

Table S18 : Exp. 3: Results on Far (15–50m) distances on LuSNAR dataset. Red indicates the best result and blue indicates the second best.

Figure S1 : Spearman rank correlation (

\rho

) between model Training Scale and evaluation metrics. The figure illustrates the correlation of the models’ total training scale with depth estimation performance (

\delta_{1}

, Abs Rel, RMSE) across standard dataset evaluations (Classic), extreme lighting conditions (Shaded), specific geological features (Semantic LuSNAR: Regolith, Rocks, Craters), and varying depth intervals (Distance: Near, Medium, Far). Statistical significance is indicated by p-values.

Figure S2 : Spearman rank correlation (p) between the percentage of Real Training Data Percentage and evaluation metrics. This figure illustrates the correlation of the proportion of real-world training data with depth estimation performance (

\delta_{1}

Figure S3 : Spearman rank correlation (p) between the percentage of Metric Supervision and evaluation metrics. This plot illustrates the correlation of the proportion of metric supervision (percentage of the training data containing metric depth) used during training with depth estimation performance (

\delta_{1}

Figure S4 : Qualitative depth estimation results on LunarSim

Figure S5 : Qualitative depth estimation results on LuSNAR

Figure S6 : Qualitative depth estimation results on Etna-LRNT

Figure S7 : Qualitative depth estimation results on Etna-S3LI

Figure S8 : Qualitative depth estimation results on Chang’e-3

Figure S9 : Qualitative depth estimation results on CHERI

Monocular Depth Estimation

Monocular depth estimation infers scene geometry and distance from a single camera image, enabling 3D perception without stereo or depth sensors.

See all 135 papers

Apr 1, 2026

MoonAnything: A Vision Benchmark with Large-Scale Lunar Supervised Data

Accurate perception of lunar surfaces is critical for modern lunar exploration missions. However, developing robust learning-based perception systems is hindered by the lack of datasets that provide both geometric and photometric supervision. Existing lunar datasets typically lack either geometric ground truth, photometric realism, illumination diversity, or large-scale coverage. In this paper, we introduce MoonAnything, a unified benchmark built on real lunar topography with physically-based rendering, providing the first comprehensive geometric and photometric supervision under diverse illumination with large scale. The benchmark comprises two complementary sub-datasets : i) LunarGeo provides stereo images with corresponding dense depth maps and camera calibration enabling 3D reconstruction and pose estimation; ii) LunarPhoto provides photorealistic images using a spatially-varying BRDF model, along with multi-illumination renderings under real solar configurations, enabling reflectance estimation and illumination-robust perception. Together, these datasets offer over 130K samples with comprehensive supervision. Beyond lunar applications, MoonAnything offers a unique setting and challenging testbed for algorithms under low-textured, high-contrast conditions and applies to other airless celestial bodies and could generalize beyond. We establish baselines using state-of-the-art methods and release the complete dataset along with generation tools to support community extension: https://github.com/clementinegrethen/MoonAnything.

Clémentine Grethen, Yuang Shi, Simone Gasparini +1

IRIT - Université de Toulouse, France · National University of Singapore, Singapore · IPAL, IRL2955, Singapore

Remote Sensing Benchmarks 3D Reconstruction

Ask PDF

Oct 20, 2026

MARE: An Open-Source Lunar Rover Simulator with Shadow Ground Truth for Perception and Autonomy

Lunar surface exploration increasingly relies on autonomous rovers that must perceive and traverse unstructured terrain under extreme illumination conditions. Near the lunar poles, low sun elevations produce long, high-contrast shadows that challenge stereo vision, terrain segmentation, and localization. Developing perception and navigation algorithms for these conditions requires large amounts of sensor data with accurate ground truth, which field campaigns and analog facilities can only provide at high cost and with limited variety. This paper presents MARE (Modular Autonomy and Rendering Environment), an open-source simulation suite for lunar rover perception and autonomy built on Unreal Engine 5 and ROS 2. MARE consists of three integrated components. Terrain Studio generates cratered and rocky lunar Digital Elevation Maps (DEMs) to build simulation scenes. A dedicated Unreal Engine project hosts two physics-based rovers with plugins for stereo, lidar, event, and Single-Photon Avalanche Diode (SPAD) cameras, streaming time-stamped sensor data and pixel-accurate ground truth. Lastly, a Guidance, Navigation, and Control (GNC) stack provides web-based mission control. Validation shows a mean Intersection over Union (IoU) of 0.79 between predicted and rendered shadows at low sun elevations, and stereo depth analysis verifies decreased pixel coverage in shadowed regions relative to sunlit terrain.

Raúl Castilla-Arquillo, Miguel Olivares-Mendez

Space Robotics (SpaceR) Research Group, SnT, University of Luxembourg, Luxembourg

Robotics Simulation Planetary Rover Robotics Space Robotics Synthetic Data Generation

Ask PDF

Apr 22, 2026

LunarDepthNet: Generation of Digital Elevation Models using Deep Learning and Monocular Satellite Images

Recent times have seen an increase in demand of high quality Digital Elevation Models (DEMs) for the lunar surface, because they are highly important for studying the moon and planning future missions. However, there is an evident lack of detailed elevation data on the Moon. To overcome this limitation, this study proposes a novel deep learning method that estimates and generates a surface elevation map directly from monocular images of the surface. The dataset used comprises of the Chandrayaan-2 Terrain Mapping Camera (TMC) images with their corresponding Digital Terrain Models (DTMs). The study proposes LunarDepthNet, which comprises of a UNet architecture to generate DEMS. It incorporates an EfficientNet encoder and custom layers to correctly learn how the light shadows on the surface relate to the actual elevation values. A combined loss function was also utilized to keep the terrain details accurate and smooth. During validation, the model showed a stable loss convergence of 12%. It achieved a mean nRMSE of 0.437 and an MAE of 4.5m in the testing stage. These results prove the model can generate dependable elevation maps from single orbital images, which are quite useful in regions of the moon where stereo-images are not available.

Aaranay Aadi, Jai Gopal Singla, Amitabh +3

School of Computer Science and Engineering Manipal University Jaipur Jaipur, India · Space Applications Centre (SAC) Indian Space Research Organisation (ISRO) Ahmedabad, India · Department of IoT and Intelligent Systems Manipal University Jaipur Jaipur, India +1

U-Net Monocular Depth Estimation Remote Sensing

Ask PDF

Abstract

Monocular Depth Estimation

Explore similar work

MoonAnything: A Vision Benchmark with Large-Scale Lunar Supervised Data

MARE: An Open-Source Lunar Rover Simulator with Shadow Ground Truth for Perception and Autonomy

LunarDepthNet: Generation of Digital Elevation Models using Deep Learning and Monocular Satellite Images