Monocular depth estimation is a fundamental capability for 3D spatial perception in robotic and intelligent systems. While data-driven discriminative methods have achieved compelling performance for monocular depth estimation, generative approaches can reconstruct fine structural details with stronger generalization. However, most existing generative diffusion models suffer from low inference efficiency caused by multi-step sampling, whereas single-step diffusion models struggle to recover high-fidelity depth maps. To mitigate these drawbacks, this paper presents FlowMoDE, a coarse-to-fine flow-matching-based framework for efficient, structure-aware monocular depth estimation supporting sim-to-real and cross-scene generalization. The core architecture first conducts coarse-level decoding conditioned on diffused features extracted from a pre-trained semantic encoder, and subsequently restores fine-grained structural patterns via expanded decoding layers. Trained exclusively on a single synthetic indoor dataset, our proposed method outperforms prior methods across five real-world indoor and outdoor datasets, yielding an average 14.1% reduction in relative error against state-of-the-art works as well as 9.9% reduction in chamfer distance for point cloud reconstruction. We further validate the generalization capability of FlowMoDE on realistic robotic scenes, with qualitative comparisons against state-of-the-art methods. The project page is https://flowmode-shao.netlify.app/
Figures & tables
Fig. 1: Comparison with existing works. Qualitative comparison in green box demonstrates the outstanding structural details from our method, compared with state-of-the-art discriminative model MoGe-2 and generative model PPD. Bubble graph with purple box denotes comparison of performance , data efficiency and computational cost of the proposed model and existing models.
Fig. 2: Pipeline of the proposed method.
NYUv2
KITTI
ETH3D
ScanNet
DIODE
Method
Year
Training Data
FLOPs (G)
AR ↓
δ1↑
AR ↓
δ1↑
AR ↓
δ1↑
AR ↓
δ1↑
AR ↓
δ1↑
Avg Rank
Data-driven Discrimitive Models
MiDaS [ 15 ]
2020
2M
184.6
11.1
88.5
23.6
63.0
18.4
75.2
12.1
84.6
33.2
71.5
19.2
DiverseDepth [ 22 ]
2021
320K
207.8
11.7
87.5
19.0
70.4
22.8
69.4
10.9
88.2
37.6
63.1
19.6
LeReS [ 23 ]
2021
354K
190.4
9.0
91.6
14.9
78.4
17.1
77.7
9.1
91.7
27.1
76.6
16.7
Omnidata [ 24 ]
2021
12.2M
195.4
7.4
94.5
14.9
83.5
16.6
77.8
7.5
93.6
33.9
74.2
16.7
TABLE I: Results on Marigold benchmarks. ∗ VAE Encoder & Decoder are not considered in original reported FLOPs calculation. # NO reported value, estimated from the description in the original paper. Best models from ALL models are in bold, while Second Best models are underlined. Best results and Second Best results from Sim-to-Real Generative Models are marked by corresponding colors.
Fig. 3: Qualitative comparison on indoor examples (first four rows) and outdoor examples (last four rows). Figures in the second row of each sample are GT and corresponding error maps based on Abs Rel.
Near-field
Far-field
NYUv2
ETH3D
Mean
Method
Acc ↓
CD ↓
Acc ↓
CD ↓
Acc ↓
CD ↓
UniDepth v2
86.7
117.2
332.4
427.1
209.6
272.2
MoGe
92.7
121.4
352.4
448.7
222.6
285.1
InfiniDepth
100.4
130.7
352.3
463.3
226.4
297.0
DepthMaster
100.0
120.9
321.8
322.3
210.9
221.6
TABLE II: Quantitative comparison of point cloud reconstruction based on monocular depth estimation. Best results and Second Best results are marked by corresponding colors.
Fig. 4: Comparison of point cloud reconstruction based on estimated depth maps.
Fig. 5: Qualitative comparison of monocular depth estimation in robotic scenarios.
Fig. 6: Qualitative comparison of point cloud reconstruction based on estimated depth maps in diverse cross-embodiment robotic scenes.
NYUv2
KITTI
ETH3D
ScanNet
DIODE
Ablation Settings
AR ↓
δ1↑
AR ↓
δ1↑
AR ↓
δ1↑
AR ↓
δ1↑
AR ↓
δ1↑
Full Model
4.0
98.1
6.2
95.9
3.9
98.8
4.3
97.6
18.8
80.4
#
Semantic Decoder
Coarse Decoder
Fine-grained Decoder
Configurations of Decoders
I
4-Layer
8-Layer
×
5.8
97.2
8.3
94.0
6.1
97.4
5.6
96.8
21.7
78.7
II
4-Layer
×
8-Layer
5.5
97.5
8.9
93.8
5.9
97.5
5.3
97.2
20.7
79.6
III
×
12-Layer
×
6.2
96.8
10.5
89.6
8.2
95.7
6.1
96.3
22.1
77.9
TABLE III: Ablation Studies on Velocity Prediction Model.
Fig. 7: Qualitative comparison of ablation settings.
NYUv2
ETH3D
Mean
Ablation
Acc↓
CD↓
Acc↓
CD↓
Acc↓
CD↓
II
118.0
119.9
455.4
383.1
286.7
251.5
IV
132.2
123.1
460.2
368.1
296.2
245.6
V
135.2
128.5
902.5
621.1
518.8
374.8
Ours
89.3
100.2
269.6
299.1
179.5
199.7
TABLE IV: Ablation results on point cloud reconstruction. Best models are in bold.
NYUv2
KITTI
ETH3D
ScanNet
DIODE
Steps
AR
δ1
AR
δ1
AR
δ1
AR
δ1
AR
δ1
1
4.4
98.1
7.5
94.7
4.6
98.5
4.5
97.6
18.8
8 0.3
2
4.0
98.1
6.2
95.9
3.9
98.8
4.3
97.6
18.8
80.4
4
4 .1
9 8.0
6 .5
9 5.6
4 .0
9 8.7
4.3
9 7.5
1 8.9
8 0.3
6
4.2
9 8.0
6.7
95.4
4.1
9 8.7
4 .4
9 7.5
19.0
80.2
8
4.2
9 8.0
6.7
95.4
4.1
9 8.7
4 .4
9 7.5
19.0
80.2
TABLE V: Ablation studies on flow matching steps. Best models are in bold, while Second Best models are underlined.
Methods
GFLOPs
ms/image
Mem (GB)
Params (M)
UniDepth v2
1925.0
23.3
3.98
353.8
MoGe
3108.0
56.8
3.65
314.2
DepthMaster
5241.5
94.7
6.71
1028.2
PPD
18158.8
193.7
4.22
804.2
FlowMoDE
3140.1
53.4
1.87
398.7
TABLE VI: Comparison of Computational Cost and Efficiency
Nanjing University of Science and Technology · State Key Laboratory of Intelligent Manufacturing of Advanced Construction Machinery · Department of Electrical and Computer Engineering, Sungkyunkwan University +2