Depth perception is fundamental for autonomous navigation, industrial inspection, and mixed-reality systems. It often operates under strict constraints on power and form factor. In such cases, monocular depth estimation (MDE) is often used. Even as recent advances continue to improve monocular depth quality, conventional frame-based MDE systems remain challenged by fast motion and high-dynamic-range scenes. This work introduces eSharp, an MDE framework that addresses these problems using an event camera. eSharp repurposes pretrained diffusion priors for event-only MDE and produces accurate metric depth with unparalleled sharpness on real data. Real-world event datasets provide only sparse LiDAR depth maps for ground-truth supervision, making dense prediction challenging. We address this limitation through cross-modal distillation, using LiDAR-calibrated dense pseudo-labels generated from a spatially co-registered frame camera to train eSharp. Experiments show that eSharp outperforms prior event-to-depth baselines in both metric-depth accuracy and spatial fidelity, establishing diffusion models as a promising foundation for event-driven 3D perception.
Figures & tables
Figure 1: eSharp produces sharp, accurate depth maps under challenging conditions. Under imaging constraints such as (a) high dynamic range, (b) sensor saturation, and (c) low light, conventional frame cameras cannot provide visual signals for reliable depth estimation. Event cameras are better suited for these challenging imaging conditions, and this work proposes eSharp, a dense metric depth estimator conditioned on monocular event cameras.
Figure 2: Overview of eSharp. The event stream is temporally binned and encoded into an event latent. An RGB-to-depth teacher generates dense pseudo-depth supervision, which is calibrated to LiDAR and encoded to depth latent. We concatenate the clean event latent and the noisy depth latent. Training proceeds in two stages: Stage I adapts the denoising U-Net for event-conditioned depth generation, while Stage II jointly fine-tunes the denoising U-Net and a lightweight calibration head to improve metric depth predictions.
Figure 3: VAE preserves binned event semantics. We feed binned event image e through a frozen VAE (without denoising). The overall event semantics are well-preserved.
Error Metrics ↓
Accuracy Metrics ↑
Dataset
Method
Abs Rel
Sq Rel
RMSE
RMSE log
SILog
δ<1.25
δ<1.252
δ<1.253
DSEC
E2Depth [ 15 ]
0.344
0.253
13.467
0.376
0.098
0.447
0.755
0.915
E2Depth ∗ [ 15 ]
0.253
0.130
10.119
0.315
0.107
0.574
0.861
0.956
EReFormer [ 28 ]
0.387
0.401
13.954
0.395
0.124
0.486
0.776
0.892
EReFormer ∗ [ 28 ]
0.286
0.208
11.369
0.325
0.109
0.569
0.839
0.944
DepthAnyEvent [ 1 ]
0.277
0.170
11.117
0.292
0.051
0.585
0.860
0.955
Table 1: Quantitative comparison on DSEC and MVSEC. We compare eSharp against open-sourced event-to-depth baselines. Results marked with ∗ are evaluated after per-image scale-and-shift alignment to the ground-truth LiDAR depth, whereas unmarked results are evaluated directly using metric depth outputs. Without test-time LiDAR calibration, eSharp achieves the best results across all DSEC metrics and leads in logarithmic errors and threshold accuracies on MVSEC. Best and second-best results are shown in bold and underline , respectively, based on unrounded values.
Figure 4: eSharp vs. eSharp-fast . eSharp (10-step diffusion) produces sharper depth predictions than eSharp-fast (1-step diffusion). The two models produce largely the same image semantics.
Figure 5: eSharp produces high-fidelity metric depth predictions on DSEC. Compared to baseline DepthAnyEvent, eSharp produces significantly sharper depth predictions, delineating cars, small obstacles on roadways, and buildings. Ω denotes results shown only at valid LiDAR pixels. All predictions are visualized using the same fixed depth mapping.
Figure 6: Metric depth predictions on MVSEC. On this dataset with lower-quality imagery (lower resolution and sparser events), eSharp still generates sharper, more accurate metric depth predictions than the baseline. Ω denotes results shown only at valid LiDAR pixels. All predictions are visualized using the same fixed depth mapping. eSharp also reconstructs the vehicle hood (curved region at the bottom).
Figure 7: Zero-shot depth estimation. We evaluated event-to-depth models trained solely on DSEC (driving scenes) on TUM-VIE (diverse egocentric scenes) to qualitatively assess domain transfer. eSharp still produces reasonable depth estimates, while DepthAnyEvent exhibits weaker transfer skills. See grayscale images (different camera pose and field of view) for visual reference.
Figure 8: eSharp predictions enable downstream segmentation. We qualitatively evaluate SAM2 [ 36 ] directly on dense depth predictions from DepthAnyEvent [ 1 ] and eSharp. The sharper scene structure preserved by eSharp yields more coherent object masks and improved segmentation quality.
Error Metrics ↓
Accuracy Metrics ↑
Dataset
Training Strategy
Abs Rel
Sq Rel
RMSE
RMSE log
SILog
δ<1.25
δ<1.252
δ<1.253
DSEC
latent optimization
0.342
0.189
6.749
0.325
0.039
0.392
0.859
0.972
[latent + metric] optimization
0.239
0.146
6.591
0.285
0.074
0.644
0.890
0.960
latent → [latent + metric] optimization
0.149
0.055
5.268
0.203
0.037
0.794
0.949
0.987
MVSEC
latent optimization
0.449
1.167
7.899
0.447
0.166
0.486
0.742
0.869
[latent + metric] optimization
0.503
2.035
9.766
0.478
0.191
0.508
0.743
0.858
Table 2: Ablation of two-stage training and metric calibration on eSharp. Lower is better for error metrics ( ↓ ), while higher is better for accuracy metrics ( ↑ ). The best result for each dataset is shown in bold .
Figure 9: Limitations. eSharp metric depth predictions depend on the quality of event streams and can lead to false-positive and false-negative object predictions.
Warping method
Image-to-depth method
MAE ↓
RMSE ↓
AbsRel ↓
δ1↑
δ2↑
δ3↑
Camera-matrix homography
DA2/LiDAR
3.033
5.683
0.120
0.870
0.963
0.986
10 m plane
DA2/LiDAR
3.057
5.721
0.120
0.868
0.963
0.986
DA3-LiDAR warp
DA3/LiDAR
2.941
5.642
0.135
0.861
0.956
0.982
DA3-metric warp
DA3-metric estimation
7.669
9.639
0.359
0.244
0.917
0.977
DA2-LiDAR warp
DA2/LiDAR
2.983
5.572
0.117
0.873
0.966
0.987
Table 3: DSEC event-space pseudo-label depth accuracy on 1000 samples after clipping depth to 0.5–80 m. Lower is better for MAE, RMSE, and AbsRel; higher is better for δ . DA stands for Depth Anything.
Method
Abs Rel ↓
Sq Rel ↓
RMSE ↓
δ1↑
δ2↑
δ3↑
Raw prediction
0.306
0.225
9.641
0.463
0.727
0.853
Depth affine LS
0.407
0.784
7.196
0.452
0.762
0.877
Inverse-depth affine LS
0.302
0.328
8.517
0.498
0.777
0.891
Log-depth affine LS
0.315
0.458
7.521
0.529
0.779
0.902
Quantile matching
0.268
0.591
6.499
0.652
0.859
0.935
Table 4: Pseudo-label alignment quality on MVSEC. We compare the raw teacher prediction, affine least-squares (LS) alignment in different spaces, and quantile matching. Metrics are computed against the available metric depth. Lower is better for error metrics ( ↓ ), while higher is better for accuracy metrics ( ↑ ). The best result for each metric is shown in bold.
Method
Stage
Trainable modules
Trainable params.
Frozen params.
eSharp
latent optimization
U-Net
865.9M
424.0M
[latent + metric] optimization
U-Net + calibration head
865.9M + 1.9K
424.0M
eSharp-fast
latent optimization
U-Net
867.6M
424.0M
[latent + metric] optimization
U-Net + calibration head
867.6M + 1.9K
424.0M
Table 5: Trainable parameter counts for latent optimization and metric optimization stages. The metric optimization stage introduces only a lightweight calibration head in addition to the event-conditioned U-Net.
Error Metrics ↓
Accuracy Metrics ↑
Dataset
Training Strategy
Abs Rel
Sq Rel
RMSE
RMSE log
SILog
δ<1.25
δ<1.252
δ<1.253
DSEC
latent optimization
0.176
0.058
5.792
0.232
0.045
0.720
0.956
0.988
[latent + metric] optimization
0.271
0.108
6.858
0.371
0.068
0.391
0.749
0.927
latent → [latent + metric] optimization
0.146
0.055
5.046
0.196
0.034
0.810
0.953
0.987
MVSEC
latent optimization
0.399
0.824
7.693
0.457
0.191
0.392
0.714
0.872
[latent + metric] optimization
0.480
1.073
7.519
0.470
0.185
0.409
0.698
0.848
Table 6: Ablation of two-stage training and metric calibration on eSharp-fast. Lower is better for error metrics ( ↓ ), while higher is better for accuracy metrics ( ↑ ). The best result for each dataset is shown in bold .
Dataset
Resolution
Method
Latency ( sframe−1 ) ↓
FPS ↑
VRAM (GB) ↓
DSEC
320×640
Depth AnyEvent [ 1 ]
0.012
83
0.29
eSharp
0.47
2.1
5.6
eSharp-fast
0.073
14
2.3
MVSEC
256×344
Depth AnyEvent [ 1 ]
0.0097
100
0.15
eSharp
0.36
2.8
5.2
eSharp-fast
0.049
20.4
2.0
Table 7: Inference efficiency on DSEC and MVSEC. Spatial (H × W) resolution of binned event data for each dataset is shown here. We report mean inference latency, throughput, and peak allocated GPU memory. Lower is better for latency and VRAM ( ↓ ); higher is better for frames per second (FPS) ( ↑ ).
Error Metrics ↓
Accuracy Metrics ↑
Method
λaff
λmetric
Abs Rel
Sq Rel
RMSE
RMSE log
SILog
δ<1.25
δ<1.252
δ<1.253
eSharp
0.0
1.0
0.262
0.132
5.973
0.266
0.035
0.598
0.912
0.976
0.3
0.7
0.161
0.066
5.341
0.212
0.039
0.782
0.944
0.984
0.5
0.5
0.166
0.072
5.390
0.214
0.038
0.777
0.940
0.983
0.7
0.3
0.149
0.055
5.268
0.203
0.037
0.794
0.949
0.987
eSharp-fast
0.0
1.0
0.146
0.055
5.031
0.196
0.034
0.809
0.953
0.987
Table 8: Ablation of calibration-loss weighting on DSEC. λaff and λmetric control the affine-parameter and metric-depth supervision, respectively. Lower is better for error metrics ( ↓ ), while higher is better for accuracy metrics ( ↑ ). Best and second-best results within each model are shown in bold and underline , respectively.
Figure
Dataset and sample identifier
Fig. 1
dsec_interlaken_00_b_rgb_000950
dsec_interlaken_01_a_rgb_000090
dsec_zurich_city_09_d_0000000094
Fig. 2
dsec_interlaken_00_f_0000000020
Fig. 3
dsec_pipeline_interlaken_00_f_0000000720
Fig. 4
dsec_zurich_city_05_b_0000000043
Table 9: Sample details used to generate the qualitative results shown in the main paper.