Performance Assessment of Neural Radiance Fields (NeRF) and

Photogrammetry for 3D Reconstruction of Man-Made and Natural

Features

Abhinav Jagan Polimera, M. M. Prakash Mohan

and K. Rajitha

Birla Intitute of Technology and Science Pilani, Hyderabad Campus, India

Keywords:

Neural Radiance Fields (NeRF), Photogrammetry, 3D Reconstruction, Ecological Modeling.

Abstract:

The present study focuses on the reconstruction of 3D models of an antenna (man-made) and a bush (natu-

ral feature) by adopting the recently developed Neural Radiance Fields (NeRF) technique of deep learning.

The performance of the NeRF was compared with the outcomes obtained by the traditional photogrammetry

methods. The ground truth geometric observation of the selected objects derived using electronic distance

measurement-based techniques revealed the efﬁcacy of NeRF compared to photogrammetry for both man-

made and natural features’ reconstruction cases. The capabilities of NeRF to reconstruct the features with

complex geometries were evident from the outcome of bush 3D reconstruction. The prospectus of canopy and

leaf level geometry estimation using NeRF will aid the enhanced modeling of vegetation-atmosphere interac-

tions. The ﬁndings presented in the study have signiﬁcant implications for diverse ﬁelds, from entertainment

to ecological modeling, and offer insights into the practical applications of NeRF in 3D reconstruction. The

outcomes of the present study attempted with a texture-less object like a bush unveiled the opportunities to

apply the NeRF techniques in precision agriculture.

1 INTRODUCTION

Three-dimensional (3D) models are versatile tools

with various applications across industries, from en-

tertainment to education, healthcare to engineering.

They enhance visualization, planning, and problem-

solving by providing an immersive and interactive

experience that traditional 2D representations cannot

match. They offer a realistic and immersive way to

visualize and interact with objects, spaces, and con-

cepts. The scientiﬁc community, especially environ-

mental researchers, considers the development of 3D

reconstruction techniques as a boon to augment the

ecological models with more structural attributes for

model calibrations (Munier-Jolain et al., 2013).

The accurate modeling of heterogeneous features

of the natural environment demanded a sophisticated

data acquisition system to generate the 3D models.

The high capital cost and processing requirements

hampered the development of 3D models in the eco-

logical domain and restricted them to 2D models,

which created more gaps from real scenarios. The

https://orcid.org/0000-0003-2484-4964

https://orcid.org/0000-0003-2269-1933

last decade’s prominent focus on climate change-

related research explored new ways of implement-

ing 3D model-derived parameters adopting the ad-

vances in computer vision techniques. In the previ-

ous two decades, digital photogrammetry techniques

revolutionized 3D topographic mapping with sufﬁ-

cient overlapping stereo-pairs(Chandler, 1999). The

requirement of sufﬁcient overlapping photos of high

resolutions attenuated the traditional photogrammetry

technique’s applications in heterogeneous feature re-

construction.

The surge in artiﬁcial intelligence and machine

learning techniques has catalyzed a revolution across

numerous domains within science and engineering.

The search for an appropriate technique that requires

a minimum number of photographs for 3D recon-

struction converged towards the recent approach of

NeRF (Neural Radiance Field). The NeRF technol-

ogy, supported by a complex neural network, en-

abled the rapid and accurate generation of 3D models

(Palestini et al., 2023).

The present study targets to utilize the capabili-

ties of NeRF for reconstructing the 3D image of a

man-made feature and a natural heterogenous fea-

tures. The research questions we address in this con-

840

Polimera, A., Mohan, M. and Rajitha, K.

Performance Assessment of Neural Radiance Fields (NeRF) and Photogrammetry for 3D Reconstruction of Man-Made and Natural Features.

DOI: 10.5220/0012396700003636

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 3, pages 840-847

ISBN: 978-989-758-680-4; ISSN: 2184-433X

text are how effectively the NeRF extracts the geo-

metrical attributes of both man-made and natural fea-

tures and how it differs from photogrammetry-derived

outcomes. Using ground-veriﬁed geometrical param-

eters estimated by accurate electronic distance mea-

surement techniques enabled the comparison between

the performances of NeRF and photogrammetry out-

comes.

2 RELATED WORK

Various global authors discussed the role of 3D data-

driven spatial models in understanding different as-

pects of ecological modeling like forest changes, re-

silience to climate change, and capacity for carbon

sequestration(Huang et al., 2019). Ecological model-

ing requires accurate canopy architecture quantiﬁca-

tion by determining the leaf area index (LAI). The re-

construction of 3D natural features like a plant canopy

geometry remains challenging due to heterogeneous

features and questions about the geometric ﬁdelity of

classical approaches (Xu et al., 2021). The research

carried out in the recent decade related to 3D recon-

struction of trees paid less attention to the estimations

of geometrical parameters like canopy crown volume,

tree height, vertical and horizontal distributions of fo-

liage and leaf angle (Gromke et al., 2015). These

geometric parameters are essential to understand the

climate-related aspects of the ecosystem.

Two commonly applied techniques for 3D re-

construction of natural features involve active range

data obtained through structured light sources such

as lasers, and the other approach utilizes overlapping

photos in conjunction with stereoscopic vision (Se-

queira et al., 1999). The leaf level traits through these

methods were less attempted due to high-cost in terms

of data volume and processing requirements.The re-

cent developments in deep-learning-based Neural Ra-

diance Fields (NeRF) that focus on synthesizing new

views of 3D objects and reconstructing 3D shapes

from a collection of images pave the way towards bet-

ter geometric estimations(Mildenhall et al., 2021).

NeRF represents a signiﬁcant shift in 3D com-

puter vision and has shown remarkable potential in

generating novel views of complex scenes (Tancik

et al., 2023). The prospect of NeRF signiﬁes a change

in research towards a more holistic understanding

and modeling of three-dimensional scenes, especially

for a natural environment. NeRF’s ability to rep-

resent detailed scene geometry with complex occlu-

sions makes it suitable for canopy architecture-related

studies which is less attempted in the present stage of

research. The fewer views requirement and effective

capture of geometric features from heterogenous en-

vironments make the NeRF technology a better option

for 3D reconstruction (Deng et al., 2022).

3 APPROACH

The overall methodology adopted for the present

study is shown in Figure 1. In the present study, we

have examined Neural Radiance Fields (NeRF) and

photogrammetry, two methods used for 3D model-

ing and reconstruction. For the comparative analysis

adopted in the study, two distinct objects were con-

sidered. The ﬁrst object, an antenna, as seen in Fig-

ure 2 (b), is predominantly composed of precise geo-

metrical shapes, while the second object chosen orig-

inates from nature, speciﬁcally, a bush (Figure 2 (a)).

The rationale behind this selection lies in the aspira-

tion to assess the effectiveness of both methodologies

in diverse contexts. It can be asserted that the pro-

cess of 3D reconstruction for geometrically ﬂawless

objects is inherently less complex when juxtaposed

with the reconstruction of natural objects, which in-

herently feature a greater degree of irregularities on

their surfaces.The assessment procedure hinges upon

the utilization of RGB (Red, Green, Blue) images de-

rived from a video source. These images, represent-

ing individual frames extracted from the video stream,

serve as the foundational input for the evaluation pro-

cess. The capabilities of NeRF and photogrammetry

techniques for 3D reconstruction are discussed fur-

ther, along with a comparison.

Photogrammetry is a versatile and widely used

technique for creating 3D models or reconstructing

objects and scenes from photographs. The overlap-

ping images captured from different viewpoints serve

as the input data for the reconstruction process. In

the initial stages of photogrammetry, distinct features

are identiﬁed and matched across overlapping images.

These features could include points, lines, or other vi-

sually distinct elements. This matching process estab-

lishes the correspondence between the same feature in

different images.

Accurate reconstruction in photogrammetry re-

quires understanding the internal and external param-

eters of the cameras used to capture the images. The

calibration of the camera’s intrinsic properties, such

as focal length and lens distortion, and determining its

position in 3D space are required for accurate 3D re-

construction. Using the calibrated camera parameters

and the correspondences established in the feature ex-

traction step, 3D points, are reconstructed through tri-

angulation. Triangulation is a mathematical process

that estimates the 3D coordinates of the features by

Performance Assessment of Neural Radiance Fields (NeRF) and Photogrammetry for 3D Reconstruction of Man-Made and Natural Features

841

Figure 1: Flowchart of methodology.

determining where lines of sight from different cam-

era positions intersect in 3D space. The reconstructed

3D points create a 3D surface or mesh.

The technical capabilities of NeRF make it a bet-

ter option for the present study as a key differentiator

compared to traditional neural networks for 3D recon-

struction. Given a set of images capturing the same

object from multiple angles along with their respec-

tive poses, NeRF learns to represent the 3D object

in a way that enables the consistent synthesis of new

views based on the training set. This instance-speciﬁc

nature allows NeRF to model and represent the subtle,

object-speciﬁc details.

The fundamental architecture of NeRF (Milden-

hall et al., 2021) involves a simple Multilayer Per-

ceptron (MLP). This MLP takes a single 5D coordi-

nate as input, comprising three dimensions for loca-

tion (x, y, z) and two for viewing direction (θ,Φ). The

output of the MLP includes the density and color at-

tributes at that spatial location. In practice, the loca-

tion is mainly used to predict density, while viewing

direction is combined with other information to pre-

dict color. Despite its simplicity, this basic architec-

ture has demonstrated the ability to perform complex

tasks and effectively capture scene geometry and ap-

pearance. A neural network is trained to estimate the

color and intensity for each point in 3D space based

on the input images and camera poses. This neural

network learns a mapping from 3D coordinates and

viewing directions to radiance values.

In NeRF, the synthesized views are obtained by

Figure 2: (a) bush (b) antenna.

querying 5D coordinates along camera rays. Clas-

sic volume rendering techniques are then employed

to project the output colors and densities into a 2D

image. The process is based on the input images and

their known camera poses, and it allows for the gen-

eration of photorealistic novel views of scenes, even

when dealing with complex geometry and diverse ap-

pearance.

Photogrammetry is a traditional method for 3D re-

construction, while NeRF is a deep learning-based ap-

proach. Photogrammetry involves capturing multiple

2D images of an object or scene from different angles,

while NeRF takes a collection of 2D images and their

corresponding camera poses as input. In photogram-

metry, a process identiﬁes common features in these

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

842

images (e.g., points or edges). It uses triangulation

techniques to determine the 3D position of these fea-

tures. At the same time, in NeRF, a neural network

is trained to model a continuous 3D scene representa-

tion by learning a function that maps 3D coordinates

to RGB values.

In practical terms, NeRF is a more recent ap-

proach that leverages deep learning to create 3D mod-

els, excelling in complex and challenging scenes, es-

pecially those with unique lighting or reﬂective prop-

erties. However, it demands signiﬁcant computational

resources during training. In contrast, while being

computationally intensive during reconstruction, pho-

togrammetry is a well-established and versatile tech-

nique suitable for a wide range of scenarios.

The evaluation of both methodologies followed

the extraction of images from a video acquired

through circumferential movement around the ob-

jects. The frame extraction rate from the video was

systematically varied, consequently altering the quan-

tity of images employed for the 3D reconstruction

process. Speciﬁcally, frame rates of 2, 3, 4, and 5

frames per second were assessed, and the resultant

images were subsequently processed through the re-

spective programs.

4 RESULTS

The results obtained after applying the photogram-

metry and NeRF are discussed in this section, along

with the corresponding reconstructed images of the

antenna and the bush. We conducted a comparative

analysis of the quantiﬁed metrics, encompassing the

number of triangles, edges, and vertices, derived from

the resulting meshes for various cases of input im-

ages. The outcomes from the 3D reconstruction for

the antenna and bush are visually represented in Fig-

ure 3, Figure 4. The numerical values of the enumer-

ation of triangles, vertices, and edges, are presented

in Table 1 for bush and Table 2 for antenna for the 3D

model obtained through NeRF.

Table 1: Number of images versus the geometric attributes

for the natural object (bush)-NeRF.

Images 76 114 152 190

Triangles 3225700 3000000 2870000 2400000

Vertices 1630000 1500000 1490000 1200000

Edges 4850000 4475000 4373000 4320000

The outcomes of the comparative analysis, which

relies on the enumeration of triangles, vertices, and

edges, are conspicuously presented in Table 3 for

Figure 3: Man-made object (a) Photogrammetery result

(b)NeRF result.

Table 2: Number of images versus the geometric attributes

for the man-made object (antenna)-NeRF.

Images 76 114 152 190

Triangles 1260000 1240000 1186000 1060000

Vertices 636000 629000 6025600 581877

Edges 1890000 1870730 1785850 1146500

bush and Table 4 for antenna for the 3D model ob-

tained through photogrammetry, in addition to be-

ing visually represented through the accompanying

graph.

Table 3: Number of images versus the geometric attributes

for the natural object (bush)-Photogrammetry.

Images 76 114 152 190

Triangles 52000 73000 86000 100000

Vertices 26000 36000 42000 58000

Edges 78000 110000 132000 140000

The outcomes of the comparative analysis, as de-

lineated in Table 1 and the accompanying graph (Fig-

Performance Assessment of Neural Radiance Fields (NeRF) and Photogrammetry for 3D Reconstruction of Man-Made and Natural Features

843

Figure 4: Natural object (a) Photogrammetery result (b)

NeRF result.

0 38

114

152

190

1.5

2.5

3.5

·10

Number of Images

Number of triangles

Bush

Antenna

Figure 5: Plot of number of images versus number

triangles-NerF.

Table 4: Number of images versus the geometric attributes

for the natural object (antenna)-Photogrammetry.

Images 76 114 152 190

Triangles 32000 38000 50000 75000

Vertices 20000 23000 33000 42000

Edges 51000 58000 70000 95000

0 38

114

152

190

0.3

0.35

0.4

0.45

0.5

0.55

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

·10

Number of Images

Number of triangles

Plot of number of images vs number triangles

Bush

Antenna

Figure 6: Plot of number of images versus number

triangles-Photogrammetry.

ure 5), elucidate notable trends. The graph pertain-

ing to NeRF (Neural Radiance Fields) reveals a dis-

cernible reduction in the number of triangles with an

increase in the number of input images. This phe-

nomenon can be attributed to NeRF’s initial assump-

tion of each data point as an independent vertex in

three-dimensional space. However, as more images

are incorporated, NeRF acquires a broader context

and amalgamates points corresponding to the same

surface. Consequently, this integration leads to a de-

crease in the count of vertices, triangles, and edges.

Conversely, the graph representing photogramme-

try(Figure 6) exhibits an opposite trend, with an aug-

mentation in its geometric complexity as more im-

ages are introduced. This phenomenon can be at-

tributed to the extraction of additional feature points

from each image, facilitating improved inter-image

matching and increasing the degree of overlap among

the acquired images. In light of the aforementioned

ﬁndings, it is evident that the acquisition of geomet-

ric attributes, such as height, area, and volume, can be

achieved even with a limited number of input images.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

844

Table 5: Comparison of the geometrical attributes estimated by NeRF and Photogrammetry.

Height (Man-made object-Antenna)

Modelled Ground Truth Error Accuracy %

NeRF 2.993 m 2.968 m 0.025 m 99.157

Photogrammetry 3.021 m 2.968 m 0.053 m 98.21

Height (Natural object-Bush)

NeRF 1.794 m 1.792 m 0.002 m 99.88

Photogrammetry 1.81 m 1.792 m 0.018 m 98.99

Max.Width (Natural object-Bush)

NeRF 0.961 m 0.96 m 0.001 m 99.89

Photogrammetry 1.046 m 0.96 m 0.116 m 87.91

Speciﬁcally, our analysis focused on determining the

height of the objects under consideration and sub-

sequently comparing these measurements to ground

truth values. To obtain precise object dimensions,

we judiciously employed control points and propor-

tionally scaled the reconstructed mesh. As a result,

the height of the antenna, as derived from the NeRF

model, was 2.993 m, while the height determined

through the Photogrammetry model was 3.021 m, and

the ground truth value was 2.968 m (Figure 7). This

comparison revealed that NeRF yields results closer

to the ground truth than photogrammetry. Similarly,

when evaluating the dimensions of the bush, the re-

spective length, breadth, and height parameters were

very close to ground truth values in the case of NeRF

compared to photogrammetry-derived results (Figure

8, Table 5). In the case of bush, the geometric values

shown in Figure 8 are the average values of measure-

ments taken in different directions across the canopy

volume.

Performing an exact computational analysis

presents signiﬁcant technical challenges due to the

fundamental differences in the underlying algorithms

employed by NeRF (Neural Radiance Fields) and

photogrammetry methods. NeRF offers the advantage

of real-time view rendering. Conversely, photogram-

metry directly produces the ﬁnalized rendered model

without the intermediate step of rendering individual

views.

Empirical observations underscore the efﬁciency

discrepancy between NeRF and photogrammetry un-

der favorable lighting conditions. NeRF demon-

strates remarkable rapidity, accomplishing approxi-

mately most of the rendering process(sufﬁcient for

extracting critical geometric features) within a few

seconds. In contrast, the output derived from pho-

togrammetry takes considerably longer, spanning sev-

eral minutes to complete the rendering process un-

der similar conditions. This discrepancy in rendering

times highlights the substantial disparity in compu-

tational efﬁciency between NeRF and photogramme-

try methodologies, particularly in scenarios character-

ized by optimal lighting conditions.

The comprehensive analysis indicates a superior

performance by NeRF across the evaluated criteria.

Furthermore, the potential applications of NeRF ex-

tend to diverse domains, including the creation of

canopy height models. An additional breakthrough

lies in the precise determination of leaf angles us-

ing NeRF, as it consistently produces highly accurate

leaf models, a feat not achieved as effectively by pho-

togrammetric reconstructions.

5 CONCLUSIONS AND FUTURE

RECOMMENDATIONS

The present study underscores the signiﬁcance of 3D

models and presents a comparative analysis of two 3D

model reconstruction techniques, namely Neural Ra-

diance Fields (NeRF) and photogrammetry. The key

ﬁndings and implications derived from the study can

be summarized as follows:

• Efﬁciency of NeRF: NeRF stands out as a highly

efﬁcient method for determining the geometric di-

mensions of objects due to its ability to achieve

accurate results with a reduced number of input

images. This efﬁciency is further emphasized

by its computational speed, making it a practical

choice for 3D reconstruction tasks.

• Accurate Complex Geometry Capture: The study

demonstrates that NeRF excels in capturing com-

plex geometries, as evidenced by the highly ac-

curate 3D reconstruction of the bush model. This

exceptional accuracy implies that NeRF holds sig-

niﬁcant potential for a wide range of applications

Performance Assessment of Neural Radiance Fields (NeRF) and Photogrammetry for 3D Reconstruction of Man-Made and Natural Features

845

Figure 7: Validation of man-made object (a) Ground Truth

(b) NeRF result (c) Photogrammetry result.

where intricate geometries need to be faithfully

represented.

• Leaf Angle Measurement: NeRF’s proﬁciency in

dealing with complex geometries extends to mea-

suring leaf angles, a task that is challenging for

photogrammetry due to the lack of sufﬁcient de-

tail in the models it generates. NeRF’s ability to

capture ﬁne details makes it suitable for applica-

Figure 8: Validation of natural object (a) Ground Truth (b)

NeRF result (c) Photogrammetry result.

tions such as leaf angle measurement, which is

critical in various contexts.

• Canopy Height Models: NeRF’s efﬁcient per-

formance in determining the geometric dimen-

sions of objects has practical implications in

the creation of canopy height models for trees.

This method allows for faster and more accurate

canopy height modeling with a reduced number of

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

846

input images, streamlining the process of assess-

ing tree canopies.

• Leaf Area Index estimations: The efﬁcacy of

NeRF revealed in the present study is directed to-

wards its potential for estimating leaf area index,

which is one of the essential climate variables.

LAI is a crucial input for the various vegetation-

atmosphere interaction models like production ef-

ﬁciency models for estimating the gross primary

productivity of vegetation. Future research in this

direction enables accurate estimation of energy

ﬂuxes, which is of prime importance in today’s

climate-changing world.

• The approach applied in the present study is

well-aligned with the requirement of the high-

throughput phenotyping (HTP) system for cash

crops like cotton, which enables phenotypic trait

estimation through these non-invasive 3D imag-

ing techniques.

REFERENCES

Chandler, J. (1999). Effective application of auto-

mated digital photogrammetry for geomorpho-

logical research. Earth surface processes and

landforms, 24(1):51–63.

Deng, K., Liu, A., Zhu, J.-Y., and Ramanan, D.

(2022). Depth-supervised nerf: Fewer views and

faster training for free. In Proceedings of the

IEEE/CVF Conference on Computer Vision and

Pattern Recognition, pages 12882–12891.

Gromke, C., Blocken, B., Janssen, W., Merema, B.,

van Hooff, T., and Timmermans, H. (2015). Cfd

analysis of transpirational cooling by vegetation:

Case study for speciﬁc meteorological condi-

tions during a heat wave in arnhem, netherlands.

Building and environment, 83:11–26.

Huang, J., Lucash, M. S., Scheller, R. M., and Klip-

pel, A. (2019). Visualizing ecological data in

virtual reality. In 2019 IEEE conference on vir-

tual reality and 3D user interfaces (VR), pages

1311–1312. IEEE.

Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron,

J. T., Ramamoorthi, R., and Ng, R. (2021). Nerf:

Representing scenes as neural radiance ﬁelds for

view synthesis. Communications of the ACM,

65(1):99–106.

Munier-Jolain, N., Guyot, S., and Colbach, N. (2013).

A 3d model for light interception in heteroge-

neous crop: weed canopies: model structure and

evaluation. Ecological modelling, 250:101–110.

Palestini, C., Meschini, A., Perticarini, M., and Basso,

A. (2023). Neural networks as an alternative to

photogrammetry. using instant nerf and volumet-

ric rendering. In Beyond Digital Representation:

Advanced Experiences in AR and AI for Cultural

Heritage and Innovative Design, pages 471–482.

Springer.

Sequeira, V., Ng, K., Wolfart, E., Gonc¸alves, J. G.,

and Hogg, D. (1999). Automated reconstruction

of 3d models from real environments. ISPRS

Journal of Photogrammetry and Remote Sens-

ing, 54(1):1–22.

Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Wang,

T., Kristoffersen, A., Austin, J., Salahi, K.,

Ahuja, A., et al. (2023). Nerfstudio: A mod-

ular framework for neural radiance ﬁeld devel-

opment. In ACM SIGGRAPH 2023 Conference

Proceedings, pages 1–12.

Xu, H., Wang, C. C., Shen, X., and Zlatanova, S.

(2021). 3d tree reconstruction in support of ur-

ban microclimate simulation: a comprehensive

literature review. Buildings, 11(9):417.

Performance Assessment of Neural Radiance Fields (NeRF) and Photogrammetry for 3D Reconstruction of Man-Made and Natural Features

847