Transparent objects are exposing one of the deepest weaknesses in conventional computer vision: the assumption that what a camera sees is a reliable map of where a surface is. A new systematic review in Applied Intelligence brings together decades of research on the non-invasive three-dimensional reconstruction of glass, plastic, liquid, and other transparent materials, outlining why these objects remain so difficult to digitize and how researchers are beginning to overcome the problem. The review, led by Xinyang Cai, Hedan Liu, Yiwen Qi, and Hao Chen, proposes a framework that connects the physics of light with the visual clues captured by cameras, the algorithms used to interpret them, and the final 3D models produced. Its message is clear: transparent-object reconstruction is no longer a niche optical puzzle, but a rapidly expanding technology frontier with consequences for robotics, manufacturing, medical imaging, and digital twins.
Most 3D scanners are designed around an apparently simple idea. A camera observes light reflected from a surface, and software uses that information to calculate depth. Structured-light scanners project patterns onto an object, while stereo systems compare corresponding points in images taken from different viewpoints. These methods work remarkably well on ordinary opaque surfaces because their appearance approximately follows the Lambertian model: light is scattered in a relatively predictable way, and the brightness of a point remains sufficiently consistent as the viewpoint changes. Transparent objects violate nearly every part of this assumption. Light can reflect from the front surface, pass through the material, bend at an interface, reflect internally, emerge from another surface, or reveal a background that is visually unrelated to the object’s actual geometry.
The difficulty becomes especially severe when an object has multiple interfaces, curved walls, liquid contents, or a highly polished finish. A ray entering glass changes direction according to Snell’s law, with the amount of bending determined by the refractive indices of air and glass and by the angle of incidence. The camera therefore does not simply observe a point on the object; it observes the result of a light path that may have crossed several surfaces. Reflections can create duplicate edges, while refraction shifts the apparent location of background features. A textured pattern projected behind a glass container may appear warped, compressed, or displaced, producing information about the object—but only if the imaging system can distinguish the deformation caused by the transparent geometry from changes caused by the background itself. This is why simply applying a standard stereo or structured-light algorithm often produces missing depth, false surfaces, or unstable point clouds.
To organize this complex landscape, the review introduces the Optical mechanism–Cue observation–Inference paradigm–Reconstruction output framework, abbreviated as OCIR. The framework begins with the optical mechanism: reflection, refraction, polarization effects, fluorescence, or background distortion. It then asks what observable cue that mechanism creates in an image. The next stage identifies the inference paradigm, such as geometric reasoning, light-path triangulation, optimization, or machine learning. Finally, it records what the method reconstructs, whether that is a silhouette-based volume, a surface mesh, a depth map, a point cloud, a neural scene representation, or a model suitable for robotic grasping. By linking these four stages, OCIR offers a way to compare methods that may look unrelated on the surface but are actually solving different parts of the same inverse problem.
One major family consists of Shape-from-X approaches, in which “X” represents an indirect visual cue. Shape-from-silhouette methods use the outer contour of an object from multiple views to build a visual hull. The approach is robust when the boundary is visible, but it generally cannot recover concavities or internal surfaces hidden from all silhouettes. Researchers have attempted to refine these reconstructions using structured light, background design, and interactive segmentation, allowing distorted patterns or subtle boundaries to provide additional geometric constraints. Polarization-based methods take a different route. When light reflects or transmits through a transparent surface, its polarization state can change in ways related to surface orientation and refractive properties. By capturing several polarization images and solving an inverse polarization model, researchers can estimate surface normals and integrate them into a 3D shape. Yet polarization methods face ambiguities, calibration requirements, and sensitivity to material properties and illumination.
Another unusual strategy uses ultraviolet fluorescence. Certain transparent materials or coatings can emit visible light after ultraviolet excitation, creating a detectable signal associated with the object’s surface rather than the background behind it. This can make otherwise invisible geometry measurable without applying a matte coating or physically modifying the object. Other Shape-from-X methods infer shape from the distortion of a known background, from defocus, from light-field changes, or from the appearance of an environment reflected or refracted through the object. These techniques demonstrate an important principle: when the object itself does not provide a stable texture, the surrounding scene can become the measurement instrument. A carefully designed background, illumination pattern, or light field can transform transparent surfaces from optical obstacles into sources of geometric evidence.
The second major category explicitly models refractive ray correspondences. Instead of treating a displaced image feature as an ordinary stereo match, these methods trace possible light paths through the object and enforce the physical constraints imposed by refraction. If a camera ray, an observed background point, and an estimated surface satisfy the relevant Snell-law relationships, the candidate geometry becomes more plausible. Light-path triangulation extends the logic of ordinary triangulation to scenes in which rays bend or reflect before reaching the sensor. Position-normal consistency methods, differentiable refraction tracing, and refractive correspondence estimation all seek to solve this problem more directly. Some systems optimize a mesh so that its predicted images match real observations, while others estimate correspondences between views by accounting for the way a transparent surface changes the apparent position of background features. The advantage is physical interpretability; the cost is computational complexity and dependence on accurate camera calibration, refractive indices, lighting assumptions, and scene modeling.
Learning-based methods have emerged as the fastest-growing branch of the field. Deep networks can estimate depth from a single RGB image, complete missing depth in RGB-D data, or combine multiple views into a coherent reconstruction. For robotic manipulation, the objective is often not a visually perfect model but a reliable estimate of the object’s location, shape, collision boundaries, and graspable regions. Systems such as ClearGrasp, DepthGrasp, RGB-D local implicit functions, Seeing Glass, and later transformer-, cascade-, diffusion-, and self-supervised approaches use color images, incomplete depth measurements, surface context, semantic features, and learned priors to fill gaps created by transparency. These models can infer plausible geometry even when direct optical measurements fail. However, plausibility is not the same as accuracy. A network trained on limited object types, backgrounds, or camera configurations may perform well in laboratory scenes and fail when confronted with unfamiliar glassware, colored plastic, strong reflections, nested containers, or natural environments.
The review also surveys neural implicit reconstruction, neural radiance fields, and 3D Gaussian splatting, technologies that have transformed multi-view scene representation. Neural radiance fields represent a scene as a continuous function that predicts density and color along camera rays, enabling novel-view synthesis through volume rendering. Researchers have adapted this idea to transparent objects by incorporating refraction tracing, self-occlusion awareness, environment matting, residual representations, polarization data, or specialized priors. Rather than producing only a conventional mesh, these methods can model how an object appears from new viewpoints and under changing illumination. Gaussian splatting offers a faster alternative by representing a scene with numerous oriented, semi-transparent 3D Gaussian primitives that can be rendered efficiently. Recent transparent-object systems use surface embeddings, physics simulation, normal constraints, de-lighting priors, residual Gaussians, and inverse rendering to improve the reconstruction of glass-like geometry. The promise is real-time performance, but representing the physical behavior of transparent materials remains difficult because visual appearance depends on view direction, background content, internal transmission, and multiple scattering events.
Datasets and evaluation standards may ultimately determine whether these advances become dependable tools or remain impressive demonstrations. The review discusses resources including ClearPose, TransCG, TransProteus, and TransPose, which provide combinations of RGB images, depth information, object masks, keypoints, multispectral observations, synthetic scenes, and grasping annotations. Real-world datasets are essential because transparent materials vary widely in thickness, tint, surface finish, internal structure, and environmental interaction. Synthetic datasets offer precise geometry and unlimited control over lighting, camera placement, refractive index, and object composition, but the gap between rendered and real optical behavior can be substantial. Evaluation may involve depth error, point-cloud accuracy, surface distance, normal consistency, intersection-over-union, mesh quality, novel-view rendering, or downstream grasp success. No single metric captures every requirement. A reconstruction that is geometrically imperfect may still support a successful robotic grasp, while a visually convincing rendering may contain serious errors in the actual surface location.
The authors identify several obstacles that will shape the next stage of research. Transparent objects are often underconstrained: many different shapes can produce similar images under particular lighting conditions. Nested surfaces, liquid-filled containers, and mixed transparent-specular materials multiply the possible light paths. Algorithms must also become more robust to uncontrolled backgrounds, motion, changing illumination, sensor noise, and objects whose refractive properties are unknown. Future systems are likely to combine physical modeling with data-driven inference rather than choosing between them. Polarization, RGB, depth, multispectral imaging, event cameras, and active illumination could provide complementary cues, while differentiable renderers may allow neural networks to learn directly from optical consistency. Large-scale benchmarks with standardized capture protocols and physically accurate ground truth could make comparisons more meaningful. In robotics, the most useful systems may be active: a robot could move a camera, alter illumination, or place a structured background to deliberately create more informative distortions.
What makes this research especially compelling is its potential to change the way machines interact with everyday materials. A robot that cannot estimate the boundary of a transparent cup may misjudge its grasp, collide with a glass panel, or mistake an empty container for open space. Industrial inspection systems may need to detect tiny defects inside transparent components without coating or cutting them. Digital twins of laboratory equipment, optical devices, packaging, and fluid systems will require models that capture not only shape but also material-dependent appearance. The systematic review presents these applications as part of a common scientific challenge: recovering hidden geometry from light that has been reflected, bent, filtered, and recombined before reaching the sensor. By framing the field through OCIR, it offers researchers a shared vocabulary—and points toward a future in which transparent objects are not invisible to machines, but understood through the optical information they were once thought to conceal.
Subject of Research: Non-invasive three-dimensional reconstruction of transparent objects
Article Title: Non-invasive 3D reconstruction techniques for transparent objects: A systematic review
Article References: Cai, X., Liu, H., Qi, Y. et al. “Non-invasive 3D reconstruction techniques for transparent objects: A systematic review.” Applied Intelligence, volume 56, article 387 (2026). https://doi.org/10.1007/s10489-026-07418-4
Image Credits: AI Generated
DOI: 10.1007/s10489-026-07418-4
Keywords: Three-dimensional reconstruction; transparent objects; non-invasive methods; depth perception; refraction; polarization imaging; neural radiance fields; 3D Gaussian splatting; robotic perception
Tags: advances in transparent object scanning technologyand medical imagingapplications of transparent object reconstruction in roboticscomputer vision challenges with transparent materialsdigital twins for transparent objectsmanufacturingnon-invasive glass and plastic digitization techniquesNoninvasive 3D reconstruction of transparent objectsoptical properties of liquids in 3D modelingovercoming transparency challenges in computer visionphysics-based light interaction in 3D imagingstructured-light and stereo 3D scanning methodssystematic review of 3D transparent object reconstruction
