Source-linked AI summary
Comparative Evaluation of 3D Reconstruction Methods for Immersive Visualization of Laboratory Objects
Brian De La Cruz, Aaron Y. Zhao, Maitrey Gramopadhye, Sawyer J. Lazar, Xianming Tan, Daniel Szafir, David S. Lawrence
TL;DR
The study asks whether current 3D reconstruction methods can create realistic holographic laboratory objects for educational use. It compares four methods through graduate-student fidelity ratings and finds that NeRF is the most consistently high-fidelity approach, especially for difficult object types.
Problem
Current reconstruction methods do not perform equally across object types, while educational research has comparatively limited evidence about their suitability for immersive learning materials.
Method
The study compares photogrammetry, NeRF, Gaussian splatting, and LiDAR by having graduate students evaluate laboratory-object holograms for shape, color, texture, and defects.
Results
NeRF produces the most consistent high-quality results, particularly for reflective objects, while shape and color are reproduced more reliably than texture.
Takeaways & Limitations
The findings provide design-relevant guidance for developing realistic immersive laboratory learning objects in AR/MR environments.
Takeaways & Limitations
Some object types remain challenging for all current reconstruction approaches, and future work will evaluate effects on comprehension, retention, and transfer.
Abstract
from arXiv · showhide
In this study, we examined whether current 3D reconstruction methods can support the creation of realistic holographic representations of laboratory objects for educational use. In this regard, we compared four approaches: photogrammetry, a neural radiance field (NeRF)-based method, Gaussian splatting, and LiDAR. These methods were used to generate holographic models of common laboratory items and their fidelity was evaluated by graduate students. Participants assessed the models for shape, color, texture, and visual defects using a repeated-measures design. Across objects, the NeRF-based method produced the most consistently high-fidelity representations, particularly for transparent, reflective, or low-texture items that were difficult to capture with other approaches. Shape and color were generally reproduced more successfully than texture, suggesting that some visual properties remain more challenging to represent accurately in educational holograms. Beyond identifying the strengths and limitations of each reconstruction method, the study demonstrates a practical workflow for creating immersive learning objects that may support pre-laboratory preparation, spatial reasoning, and student engagement in AR/MR-based educational environments. These findings offer design-relevant insights for educators and researchers developing immersive digital learning experiences.
1. Introduction
AR/MR can support spatial and conceptual learning, but its educational value depends on realistic digital objects. This study addresses limited educational comparison of reconstruction methods by evaluating four approaches for laboratory holograms.
- Educational context: Immersive technologies may support engagement, motivation, and conceptual understanding when designed with clear pedagogical goals.
- Educational context: AR/MR lets learners inspect objects from multiple viewpoints and connect abstract representations to physical experience.
- Motivation: Poorly rendered virtual objects can reduce learner trust, limit engagement, and weaken instructional value by failing to preserve shape or texture.
- Research problem: Reconstruction methods vary in suitability across geometric, reflective, transparent, and low-texture objects, limiting the usefulness of methods that work well for only one object type.
- Research problem: Educational research has comparatively little evidence on how current reconstruction methods affect the usability and educational potential of AR/MR learning objects across varied materials.
- Study approach: The study compares photogrammetry, a NeRF-based method, Gaussian splatting, and LiDAR using laboratory objects evaluated by graduate students for shape, color, texture, and defects.
- Study contribution: The study aims to identify methods that can produce realistic digital objects for immersive learning environments, beyond comparing technical reconstruction quality alone.
2. Materials and Methods
The study reconstructed twelve laboratory objects with four imaging technologies, standardized their AR presentation, and evaluated fidelity through repeated within-subject assessments.
- Objects and capture: Twelve laboratory objects were selected to vary in transparency, reflectivity, texture, and geometric complexity.
- Imaging methods: Objects were scanned using four technologies: photogrammetry, Gaussian splatting, LiDAR, and a NeRF-based method.
- Imaging methods: Photogrammetry uses photographs to construct sparse 3D data emphasizing vertices, edges, and faces, but can struggle with reflective or low-texture objects.
- Imaging methods: Gaussian splatting projects millions of point-based Gaussian distributions and can interpolate missing information, potentially introducing geometric inaccuracies.
- Imaging methods: LiDAR calculates distances from laser-pulse return times, producing geometry-focused point clouds that are less effective for color and fine texture.
- Imaging methods: NeRF maps light, color, and density as functions of position and viewing angle through neural-network training, but requires high computational demand for large scenes.
- AR presentation: Models were formatted in Unity for HoloLens viewing with consistent orientation, and a PC-driven setup supplied processing for the stream-heavy application.
3. Results
Across laboratory objects, reconstruction quality depended on object characteristics and method. NeRF often produced stronger results for difficult reflective or transparent objects, while some methods performed well on simpler items.
- Object-level findings: For the acetone wash bottle, all methods reproduced color well, while LiDAR and photogrammetry struggled with shape and photogrammetry produced texture problems and defects.
- Statistical analysis: Statistical analysis for the wash bottle found method differences for shape, texture, and defects, but not color.
- Statistical analysis: Gaussian splatting and NeRF produced higher-fidelity wash-bottle models than photogrammetry and LiDAR, particularly regarding fewer defects.
- Object-level findings: For the treated glass filter flask, photogrammetry and LiDAR failed to generate effective models, while NeRF produced the only structurally well-defined model.
3.3. Object 3. Heating Mantle.
The heating mantle was reproduced successfully across all four reconstruction methods, indicating that this object was comparatively easy to model.
- 3.3. Object 3. Heating Mantle.: All four reconstruction methods furnished excellent holographic models of the heating mantle.
3.6. Object 6. Laboratory Jack.
All four scanning methods produced excellent holographic models of the laboratory jack. Its relatively simple, low-reflectivity appearance supported strong reproduction, although texture and color can be slightly diminished for complex surfaces.
- The jack has a smooth medium-blue plate, slightly rusted metal supports, and overall low reflectivity.
- All four scanning methods produced excellent 3D models of the blue laboratory jack.
- The highly textured cork ring was well reproduced by all methods, with little or no defects.
- Texture and color capture for the cork ring fell slightly below the best results obtained for other objects.This suggests that high surface complexity remains difficult to reproduce precisely.
- The LiDAR model of the cork ring was rated best, but significant defects remained visible.
- No method produced a high-fidelity model of the semi-transparent, moderately reflective plastic beaker.LiDAR and NeRF produced the best, although modest, results.
3.9. Object 9. Bunsen Burner.
The Bunsen burner was reproduced well by all scanning methods, despite its non-transparent metal surface. In contrast, the mortar and pestle models were generally poor, with LiDAR outperforming Gaussian splatting but still showing deformities.
- All methods generated good Bunsen burner models with relatively few defects.The object is non-transparent and has a dull, non-homogenous finish.
- Texture capture for the Bunsen burner was somewhat weaker than the best results for other objects.The reduction may reflect challenges in reproducing complex surfaces.
- All scanning methods produced excellent holographic replicas of the structurally and color-complex object examined next.Its components included Styrofoam, aluminum foil, colored caps, and differently oriented geometries.
- LiDAR was the only method that performed adequately for the smooth, highly reflective mortar and pestle object.
- Gaussian splatting was by far the worst performer for the reflective mortar and pestle object.
- Only subtle quality differences appeared between models of the unsprayed and cyclododecane-sprayed objects.The spray was used to create a temporary matte finish that reduced reflectivity.
3.12. Method Performance.
NeRF produced the most consistently high-fidelity holographic models across the 12 objects, while performance varied substantially with object properties. Shape and color were generally easier to reproduce than texture, and transparent or reflective objects remained challenging.
- Across 12 objects, NeRF produced the most consistently high-fidelity holographic models.
- NeRF produced the most models above the high-fidelity threshold of ≥ 4 and the fewest at or below the fair-to-poor threshold of ≤ 2.5.
- LiDAR performed well for selected objects, whereas photogrammetry and Gaussian splatting more often produced lower-rated models.
- Methods performed best for objects with clear structural features and low visual complexity.Examples included the heating mantle, laboratory jack, Bunsen burner, and test tube rack.
- Transparent, reflective, or smooth featureless objects were more difficult to reproduce accurately.Examples included the glass filter flask, amber bottle, plastic beaker, and mortar and pestle.
- Shape and color were generally reproduced more successfully than texture across methods and objects.Texture was the most difficult attribute to capture consistently, and defects were more common for transparent or reflective objects.
3.15. Statistical Comparisons.
Repeated-measures analyses found significant differences among reconstruction methods for several object–attribute combinations, especially shape, texture, and defects. NeRF and sometimes LiDAR outperformed alternatives on challenging objects, while simpler objects showed comparable results across methods.
- Statistical Comparisons: Repeated-measures analyses confirmed significant method differences for several object–attribute combinations.The strongest differences involved shape, texture, and defects.
- Statistical Comparisons: NeRF and, in some cases, LiDAR outperformed photogrammetry and Gaussian splatting for challenging objects.
- Statistical Comparisons: For several simpler objects, all four methods produced comparably strong results.
- Educational Relevance: Participants reported that the holographic models were easy to view and assess.
- Educational Relevance: The results indicate that realistic AR/MR representations can support pre-laboratory instruction, but fidelity depends strongly on object properties.This setting requires students to recognize equipment, anticipate spatial relationships, and prepare for safe laboratory work.
- Educational Relevance: The workflow creates realistic 3D digital surrogates of laboratory objects that can be inspected from multiple viewpoints.
- Educational Relevance: The workflow may strengthen safety preparation and spatial understanding in AR/MR-based learning environments.
4. Discussion
Immersive laboratory representations are feasible with existing tools, but fidelity varies by object properties and visual attribute. The findings support matching reconstruction methods and instructional goals rather than assuming one method suits every object.
- Reconstruction fidelity: The NeRF-based method produced the most consistent results across object types and may be especially useful for reflective, transparent, or low-texture items.These object types were difficult for other approaches to capture accurately.
- Reconstruction fidelity: Shape and color were reproduced more reliably than texture, indicating that texture remains a challenging visual property for educational holograms.The discussion links texture more closely to subtle realism and material perception than to basic object recognition.
- Method selection: Simple, low-reflectivity objects were generally reproduced well by all four methods, whereas transparent or highly reflective objects remained challenging.Method suitability therefore depends strongly on the visual and structural properties of the scanned object.
- Method selection: Educators and designers should match reconstruction methods to instructional purpose and object type, using more capable methods for complex or reflective objects when needed.Simpler methods may suffice for less demanding instructional tasks.
- Educational use: The workflow creates holographic learning objects that can be inspected from multiple viewpoints in a standardized, intuitive format.The resulting objects may support pre-class preparation, guided exploration, active learning, spatial reasoning, and procedural understanding.
- Educational use: Students responded positively to the holographic images, with high agreement regarding realism and frequent expressions of surprise and enthusiasm.The representations may support observational learning, spatial reasoning, conceptual transfer, engagement, and preparation for hands-on laboratory work.
- Scope and future work: Current limitations with highly transparent or reflective objects define a design threshold, while the technology remains suited to simulating many familiar laboratory objects.Future work will evaluate comprehension, retention, transfer, and training across learner populations and instructional settings.
5. Conclusions
Current 3D reconstruction methods can create realistic AR/MR learning objects, with NeRF producing the most consistent quality across laboratory objects. However, texture and some object types remain challenging, and the study did not directly measure learning outcomes.
- NeRF-based technology produced the most consistent high-quality holographic results, especially for reflective, transparent, or difficult-to-capture objects.
- Shape and color were more readily reproduced than texture, while some object types remained difficult for all tested approaches.
- The demonstrated workflow generates AR/MR learning materials that can be explored from multiple perspectives.
- These holographic objects may support spatial understanding, object recognition, and pre-laboratory preparation.
- The study provides design-relevant evidence for immersive educational resources but did not directly measure learning outcomes.
- Future research should examine learning, engagement, cognitive load, and method suitability across educational goals and object types.
Supplementary Material
The supplementary material documents the laboratory objects, model-quality assessments, and statistical comparisons used to evaluate the four reconstruction methods. It includes object-specific results across shape, texture, color, and defects, with post hoc method comparisons for statistically meaningful differences.
- Supplementary tables list twelve common laboratory objects used in the reconstruction evaluation.
- The supplementary materials cover object-specific assessments of shape, texture, color, and defects for photogrammetry, Gaussian splatting, LiDAR, and NeRF.
- Friedman analyses identify statistically meaningful differences for each object and visual attribute among the four methods.
- Post hoc paired comparisons report p-values for method pairs on objects and attributes showing statistically significant differences.
- For object 1 shape, Photogrammetry–NeRF comparisons yielded p = 0.0040, while Gaussian–NeRF comparisons yielded p = 0.11.
- For object 8 color, Photogrammetry–Gaussian yielded p = 0.044, whereas Gaussian–NeRF yielded p = 0.82.