Monocular depth estimation remains challenging for transparent objects, where refraction and transmission are difficult to model and break the appearance assumptions used by depth networks. As a result, state-of-the-art estimators often produce unstable or incorrect depth predictions for transparent materials. We propose SeeClear, a novel framework that converts transparent objects into generative opaque images, enabling stable monocular depth estimation for transparent objects. Given an input image, we first localize transparent regions and transform their refractive appearance into geometrically consistent opaque shapes using a diffusion-based generative opacification module. The processed image is then fed into an off-the-shelf monocular depth estimator without retraining or architectural changes. To train the opacification model, we construct SeeClear-396k, a synthetic dataset containing 396k paired transparent-opaque renderings. Experiments on both synthetic and real-world datasets show that SeeClear significantly improves depth estimation for transparent objects. Project page: https://heyumeng.com/SeeClear-web/
Figures & tables
Figure 2: Pipeline Overview. Starting from an image, we first apply a segmentation model to obtain the transparent object mask. Guided by the mask and the image, a latent diffusion model generates an opacified image of the transparent object. A mask refinement module then predicts a soft blending mask to alpha-composite the generated opaque region with the original background, producing the final composited image. The composited image is finally fed into a depth model to estimate accurate depth.
Figure 3: Rendering Pipeline. We build the SeeClear-396k dataset for transparent-object depth estimation. For each object–scene configuration, Blender renders paired transparent ( Itr ) and opaque ( Iop ) images with identical geometry, camera pose, and illumination. Viewpoint, lighting, and anisotropic shape variations are systematically sampled to produce diverse training data with aligned depth, normals, and masks.
Figure 4: Qualitative Comparison. We evaluate transparent-object depth estimation on ClearGrasp [ 42 ] (Columns 1–3) and TransPhy3D [ 59 ] datasets (Columns 4-6). Compared with the baseline. SeeClear produces accurate transparent-object depth. Depth maps in the circle are normalized in grayscale.
Region
Method
AbsRel ↓
SiLog ↓
RMSE ↓
iRMSE ↓
MAE ↓
δ1.025↑
δ1.05↑
δ1.10↑
ToM
Depth4ToM [ 9 ]
0.040
0.034
24.83
0.009
20.94
41.41
70.48
92.56
MODEST [ 27 ]
0.062
0.041
39.31
0.013
33.09
29.22
54.01
79.60
D4RD [ 51 ]
0.069
0.031
39.95
0.014
36.18
16.21
42.70
78.25
DKT † [ 59 ]
0.035
0.024
22.43
0.008
18.59
52.69
81.29
94.04
Marigold [ 20 ]
0.047
0.036
28.10
0.011
24.31
38.58
64.49
88.10
GenPercept [ 57 ]
0.039
0.028
24.09
0.009
20.30
47.79
75.21
92.71
Table 1: ClearGrasp Real. Comparison on the real-world test set across three evaluation regions. ToM : transparent-object region. All : full image. Other : non-transparent region. † : fine-tuned on this dataset. x : best result.
Table 5
Figure 5: In-the-Wild Qualitative Comparison (Part I). Scenes include a pair of glasses (row 1), multiple transparent objects with liquid and inter-object occlusion (row 2), a transparent container with an out-of-distribution shape enclosing opaque chocolate objects under a plastic film (row 3), a transparent plastic drawer cabinet containing opaque objects (row 4), multiple transparent objects with inter-object occlusion and liquid contents (row 5), a single transparent glass containing liquid and plants extending from inside to outside (row 6), and multiple transparent objects with opaque labels and liquid contents (row 7). Depth4ToM produces predictions with blurrier boundaries compared with other methods; in rows 1, 5, and 6 the transparent surfaces appear translucent, and in rows 2 and 7 the depth fails to reflect object surface geometry such as cup openings and bottle shapes. Depth Anything V3 and MoGe-2 exhibit pronounced translucency across all rows. Marigold and GenPercept both show translucency in rows 1, 3, 5, 6, and 7; additionally, Marigold exhibits translucency in row 4, and GenPercept exhibits translucency in row 2. SeeClear produces no translucency across all scenes, accurately recovers object surface geometry, and preserves seamless depth transitions at material boundaries.
Figure 6: In-the-Wild Qualitative Comparison (Part II). Scenes include a transparent storage box containing opaque objects (row 1), multiple transparent objects with liquid and inter-object occlusion (row 2), objects with mixed transparent and opaque materials (row 3), a single transparent object (row 4), multiple transparent objects with opaque labels and liquid contents (row 5), multiple transparent objects (row 6), and an out-of-distribution scene combining transparent and opaque materials (row 7). Depth4ToM produces blurry boundaries throughout; in rows 2–6 the depth fails to reflect object surface geometry such as cup openings, kettle shapes, and the surface shape of bottles and containers, and in row 7 transparent objects show no discernible shape in the depth map. Depth Anything V3 and MoGe-2 exhibit translucency across all rows. Marigold and GenPercept both exhibit translucency in rows 1, 3, and 7; additionally, GenPercept shows translucency in row 4. Marigold produces uneven depth transitions in rows 2, 3, and 5, and GenPercept exhibits similar unevenness in rows 3 and 5. In rows 4 and 6, Marigold produces depth maps that do not match object geometry (e.g., missing cup openings), and GenPercept shows the same issue in row 6. SeeClear accurately recovers object geometry including openings and surface contours, produces consistent depth under occlusion, and maintains seamless transitions at material boundaries.
Figure 8Figure 9
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Figure 12: Qualitative Comparison on TDoF20. DKT produces inaccurate object boundaries within the transparent-object region and occasionally fails to recover the object structure entirely (dashed circle). Our method produces depth predictions consistent with the ground truth across all five scenes. The inset grayscale images show depth normalized within the transparent-object mask region; the intensity reflects the depth values within the mask and does not imply the presence of a solid object at that location. Depth maps in the circle are normalized in grayscale.
Transparent objects pose a fundamental challenge for depth estimation and 3D reconstruction due to their violation of Lambertian assumptions, leading to severe geometry degradation in downstream tasks. To address this, we propose a novel geometry-guided preprocessing framework GHOST that leverages visual foundation models to transform transparent regions into opaque, structurally consistent representations without requiring downstream model retraining. Specifically, our pipeline utilizes (1) TransDINO and (2) TransDecomp to disentangle masks and transparency physical properties, while (3) DAF-Net recovers surface normal priors to encode geometric curvature. Subsequently, (4) GeoSemTransNet integrates these multi-modal cues to synthesize a texture-rich opaque RGB image that preserves the transparent object's 3D structure. Extensive experiments demonstrate that our method significantly enhances the accuracy of state-of-the-art depth estimation and reconstruction models on transparent objects by restoring essential photometric cues.
Langxu Zhao, Zuan Gu, Tianhan Gao
School of Software, Northeastern University, Shenyang, China.
Transparent objects are common in daily life, and it is important to understand their multilayer depth, including the transparent surface and the objects behind it. Existing methods for multilayer depth typically extend single-layer prediction. They define layers by the front-to-back ordering of 3D points and predict the layers sequentially. However, as layered geometry can admit multiple valid groupings of 3D points into layers, a predefined grouping strategy is inherently restrictive. In this work, we propose SeeGroup, a multi-layer depth estimation method that avoids imposing a predefined grouping and allows the model itself to adaptively assign surfaces to depth maps. We formulate per-pixel multi-layer depth as a point process, treating depth layers as unordered events along each camera ray. This induces a permutation-invariant likelihood over the observed depth layers, yielding a loss that naturally supports arbitrary layer groupings. Experiments demonstrate that our method significantly advances the state of the art of multi-layer depth estimation, improving quadruplet relative depth accuracy on LayeredDepth benchmark from 61.34% to 70.09%. Code is available at https://github.com/princeton-vl/SeeGroup.
Hongyu Wen, Jia Deng
Department of Computer Science, Princeton University
Standard depth sensors systematically fail on transparent surfaces, creating corrupted 3D maps and severe navigation hazards. While specialized hardware sensors can detect glass, they lack modularity and have extensive hardware dependencies. Consequently, learning-based monocular depth estimation has emerged as a compelling alternative. However, domain-specific glass-aware monocular depth estimators struggle with unfamiliar indoor layouts; restricted by the severe scarcity of real-world glass depth annotations, they fail to generalize zero-shot to new settings. This motivates us to explore whether the extensive priors of text-to-image diffusion models can enable generalizable perception of transparent surfaces. We introduce SILICA, a unified pipeline leveraging these priors to jointly predict glass segmentation and glass-aware depth. This mutual information exchange establishes a robust spatial hierarchy, entirely eliminating the need for paired real-world glass depth annotations. Subsequently, we use the predicted segmentation mask to explicitly filter incorrect glass depth points from standard sensors, recovering accurate metric glass depth for downstream 3D mapping and autonomous collision avoidance. Supported by our novel Mirage 18k dataset, extensive experiments demonstrate that SILICA achieves remarkable zero-shot transfer across diverse, unseen environments, outperforming state-of-the-art models by almost 20% and setting a new benchmark for transparent surface perception.