Lightweight Transformer Occupancy Networks for 3D Virtual Object
Reconstruction
Claudia Melis Tonti
a
and Irene Amerini
b
Sapienza, University of Rome, Italy
Keywords:
Mesh, Generation, Point Cloud, Computer Vision, 3D.
Abstract:
The increasing demand for edge devices highlights the necessity for modern technologies to be adaptable to
general-purpose hardware. Specifically, in fields like augmented reality, virtual reality, and computer graphics,
reconstructing 3D objects from sparse point clouds is highly computationally intensive, presenting challenges
for execution on embedded devices. In previous works, the speed of 3D mesh generation has been prioritized
with respect to preserving a high level of detail. Our focus in this work is to enhance the speed of the inference
in order to get closer to real-time mesh generation.
1 INTRODUCTION
In recent years, the growing importance of computer
vision in technologies like autonomous navigation
(Urmson et al., 2008), robotic mapping (Zhu et al.,
2017), augmented reality (AR) (Alhaija et al., 2017),
and gaming has increased the demand for efficient
processing of multidimensional data such as images
and videos.
Traditional methods for 3D reconstruction, such
as Signed Distance Functions (SDFs), require heavy
computations and slow processes for mesh genera-
tion. Newer approaches like ConvONet (Peng et al.,
2020) achieve faster and more efficient 3D recon-
struction compared to earlier methods like DeepSDF
(Park et al., 2019). However, real-time reconstruc-
tion from point clouds remains underexplored, despite
its importance for applications on low-performance
hardware, such as AR or immersive reality.
Point clouds, created using LiDAR sensors or esti-
mated via Monocular Depth Estimation (MDE) (Papa
et al., 2023), act as lightweight 3D mesh compres-
sions. Their sparsity can be adjusted to balance detail
retention and computational efficiency, allowing on-
demand reconstruction of 3D objects within specific
areas, such as those observed in AR viewers.
Our work focuses on efficient 3D object recon-
struction for AR applications, with use cases includ-
ing interior design, game development, and virtual ar-
a
https://orcid.org/0009-0002-0574-612X
b
https://orcid.org/0000-0002-6461-1391
tifact reconstruction (Patil et al., 2023; Virtanen et al.,
2020). We leverage lightweight implicit representa-
tions based on Convolutional Occupancy Networks
(ConvONet) (Tonti et al., 2024), extracting meshes
from point clouds to prioritize speed over accuracy.
This ensures smooth interaction in AR environments,
avoiding delays that could disrupt the user experience.
To enhance ConvONet, we integrate an efficient
transformer, capitalizing on its ability to process com-
plex spatial relationships effectively. Our method is
validated on the ShapeNet benchmark dataset.
The paper is structured as follows: Section 2 re-
views related work, Section 3 introduces the proposed
ConvONet with a Dynamic Vision Transformer, Sec-
tion 4 provides implementation details, Section 5
presents experimental results, and Section 6 con-
cludes with future directions.
2 RELATED WORKS
This section reviews key methods for representing
and reconstructing 3D shapes, focusing on various
representation techniques. Early approaches, such
as Signed Distance Functions (SDF) and Deep SDF
(Mescheder et al., 2019; Park et al., 2019), describe
surfaces as a continuous function that assigns a signed
distance value to every point in 3D space. While ef-
fective for defining geometry, SDFs require signifi-
cant memory and cannot infer surfaces in unobserved
regions, leading to gaps in the reconstructed 3D rep-
408
Tonti, C. M. and Amerini, I.
Lightweight Transformer Occupancy Networks for 3D Virtual Object Reconstruction.
DOI: 10.5220/0013377400003912
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 1: GRAPP, HUCAPP
and IVAPP, pages 408-414
ISBN: 978-989-758-728-3; ISSN: 2184-4321
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

Figure 1: Graphical representation of the proposed ConvONet pipeline. The ConvONet pipeline begins with an input point
cloud containing N points. In the Encoder, the ResNet PointNet (Qi et al., 2017) individually encodes these points, while a
plane predictor network globally encodes the entire point cloud to learn L dynamic planes and their specific features. These
plane features are combined with the per-point features and projected onto the learned dynamic planes. The dynamic planes
are summed and overlapped between each other, then processed in the Dynamic Vision Transformer. In the decoder, for
a given query point p ∈ R
3
, the feature vector ψ(x, p), where x is the input, is derived from bilinear interpolation of the
processed feature planes. Finally, a fully-connected network predicts the occupancy probability f
θ
(p, ψ(p, x)) at point p.
resentation (Curless and Levoy, 1996; Choy et al.,
2016).
Deep Local Shapes (Chabra et al., 2020) address
these limitations by using neural networks to ap-
proximate SDFs in small regions, improving infer-
ence efficiency by breaking down scenes into inde-
pendent local features. Implicit function-based meth-
ods (Chibane et al., 2020) further advance this idea by
representing 3D objects with multi-scale 3D vectors
that encode both local and global structures, enabling
better handling of complex geometries.
Genova et al. (Genova et al., 2020) introduced Lo-
cal Deep Implicit Functions (LDIF), a hybrid repre-
sentation that combines local implicit functions with
structured templates to decompose space efficiently.
LDIF uses a latent vector representation for finer de-
tail and leverages PointNet (Qi et al., 2017) to encode
both local and global features for 3D reconstruction
tasks. These representations focus on breaking down
large-scale geometries into manageable components
while retaining critical surface details.
ConvONet (Peng et al., 2020) proposed a contin-
uous implicit grid representation that combines con-
volutional layers and occupancy networks to estimate
the probability of a point belonging to the surface of
the object. This hierarchical encoding of 3D features
enables accurate reconstructions and inspired meth-
ods like DPConvONet (Lionar et al., 2021a) and Neu-
ralblox (Lionar et al., 2021b). DPConvONet projects
point cloud features onto dynamic planes, capturing
surface normals for unoriented inputs, while Neural-
blox supports nearly real-time 3D scene generation
using occupancy grids from point clouds.
Despite these advancements, achieving real-time
reconstruction remains challenging. Neuralblox of-
fers faster mesh generation but sacrifices detail, and
Lightweight ConvONet (Tonti et al., 2024) attempts
to optimize efficiency through architectural compres-
sion, yet its inference time is still not real-time.
To address these challenges, we propose integrating
Dynamic Vision Transformers (Dynamic ViT) (Rao
et al., 2023) into Lightweight ConvONet. This inte-
gration replaces the bottleneck U-Nets, enabling more
efficient processing of point cloud projections while
leveraging the transformer’s ability to handle complex
spatial relationships. This approach aims to improve
both speed and accuracy, moving closer to real-time
performance.
3 PROPOSED METHOD
Our research aims to increase inference speed to gen-
erate fast and high-quality 3D meshes. In this work,
we analyze the effects of integrating an efficient trans-
former architecture into the ConvONet pipeline in or-
der to decrease the generation time.
The rest of this section is organized as follows:
Section 3.1 introduces the structure of the network,
and Section 3.1.3 presents the architecture of the Dy-
namic Vision Transformer.
3.1 The Proposed Network
The architecture of the proposed network, shown in
Figure 1, has an encoder-decoder structure.
The pipeline, as shown in the figure, begins with
a sparse point cloud input. The encoder, detailed in
Section 3.1.1, processes this input to extract per-point
and global features, which are then used to estimate
the dynamic planes. These planes simplify the envi-
ronment by transforming it from 3D to 2D, reducing
computational complexity during encoding.
Once the points are projected onto these planes,
the overlapping projections are processed by the Dy-
namic ViT, as described in Section 3.1.3. In the de-
coder, outlined in Section 3.1.2, the planar features
are utilized to perform bilinear interpolation for each
Lightweight Transformer Occupancy Networks for 3D Virtual Object Reconstruction
409

point in 3D space, calculating its occupancy probabil-
ity. This approach balances efficiency and precision
in 3D reconstruction.
3.1.1 Encoder
The encoder-decoder network is based on the ground-
work of Lionar et al. The network processes a point
cloud input with a ResNet-based PointNet. The
encoder extracts per-point features using a ResNet
PointNet and learns a set of dynamic planes with a
plane predictor network. The ResNet PointNet con-
sists of five ResNet blocks, a PointNet, and a fi-
nal ResNet block, encoding the point cloud point-
wise. The plane predictor network uses a PointNet for
global max pooling to predict the set of planes. These
features are converted to plane parameters, defining
the plane by its normal vector (a, b, c).
The per-point and planar features are summed and
projected onto the predicted dynamic planes using a
basis change and orthographic projection, followed
by normalization as proposed by Lionar et al. Every
plane with projection is summed up with the others
and then processed with Dynamic Vit (DyVit), pre-
sented in Section 3.1.3. The outputs of the trans-
former are the planar features that will be processed
in the Decoder.
3.1.2 Decoder
Given the estimated planar features, the decoder pre-
dicts the probability of each point in 3D space being
within the occupancy grid. This probability is derived
from the sum of the bilinear interpolations of the es-
timated planar features. The network is based on a
fully connected network of five ResNet blocks, with
input features summed to the feature vector in each
block. After estimating the occupancy probability of
each point, the point cloud is turned into a mesh.
3.1.3 Dynamic Vision Transformer
Compared to Convolutional Neural Networks
(CNNs), Vision Transformers (ViTs) (Touvron et al.,
2022) are better at capturing long-range relationships
between pixels, enabling them to process images with
more comprehensive detail. The main components of
a ViT are:
• Patch Embedding: The input image is divided into
smaller patches, which are flattened and projected
into a fixed-dimensional space. Positional embed-
dings are added to retain spatial information for
each patch.
• Multi-head Self-Attention (MSA): This mecha-
nism allows the model to understand global rela-
tionships by enabling interactions between differ-
ent patches (tokens). MSA calculates query (Q),
key (K), and value (V) matrices, using softmax to
create an attention matrix that captures dependen-
cies across the input. Parallel attention layers en-
hance the model’s ability to capture complex fea-
tures. It computes query (Q), key (K), and value
(V) matrices, transforming scores into probabil-
ities with softmax, thereby forming an attention
matrix that aggregates global information across
the sequence.
• Feed-Forward Neural Network (FNN): After the
self-attention step, the output is normalized and
processed by a feed-forward network to refine the
features further.
Rao et al. (Rao et al., 2023) proposed DyVit
which dynamically removes less important tokens
from transformer blocks based on their contribution to
the output. During training, the threshold for selecting
important tokens is adjusted automatically, and only
the most relevant tokens are retained during inference.
This reduces computational costs while maintaining
accuracy. To handle performance drops caused by to-
ken pruning, DyVit uses knowledge distillation (KD).
A teacher model with all tokens intact guides the
pruned student model, helping it learn rich represen-
tations and retain performance.
The structure of DyViT adapted for our task is pre-
sented in Figure 2. The first part of the student and
teacher networks is almost identical: the input, which
consists of the 3 overlapped projection planes, is di-
vided into patches and processed by the patch embed-
der, then, the processed patches are transmitted to the
self-attention mechanism. The student model, differ-
ently from the teacher model, processes the patches
with self-attention. The attention output is processed
with the prediction module which is in between the
transformer blocks to delete less informative tokens
selected by the features given by the previous layer.
In this way, fewer tokens are processed in the follow-
ing layers. This technique is called sparsification and
the sparse tokens are processed in the self-attention
which returns the output of the model, which is the
resulting planar features.
To train DyVit effectively, the total loss is a com-
bination of three components:
• Cross-Entropy Loss Rao et al. use the standard
cross-entropy to represent the distance between
the ground truth and the result of DyViT. In our
case, we don’t have a ground truth for the projec-
tion planes so, we represent instead the distance
between the DyVit result and the plane projections
given as input. We took this decision because we
hypothesize that the nearest points to the input
GRAPP 2025 - 20th International Conference on Computer Graphics Theory and Applications
410

Figure 2: Graphical representation of the student-teacher architecture of Dynamic Vision Transformer with an input of 3
overlapped projection planes.
point cloud points have the highest probability of
lying on the surface of the mesh so, this distance
should be minimized. The loss is defined in this
shape:
L
ce
= CrossEntropy(y, y) (1)
where y is the input patch of the overlapped pro-
jection layers and y is the prediction of DyViT.
• Self-Distillation loss This loss is used to mini-
mize the influence on performance caused by to-
ken sparsification. To do so, the final remaining
tokens of DyViT should be as close as possible to
the tokens of the teacher model. This minimiza-
tion is represented by the following equation:
L
distill
=
1
∑
B
b=1
∑
N
i=1
ˆ
D
b,S
i
∑
b=1
B
∑
i=1
N
ˆ
D
b,S
i
(t
i
− t
′
i
)
(2)
where t
i
and t
′
i
are the resulting token of DyVit
and the teacher model respectively,
ˆ
D
b,S
i
is the de-
cision mask for the b-th step and the s-th sparsifi-
cation stage.
• KL Divergence with this loss it is minimized the
difference between the results of DyViT and the
results of the teacher using KL divergence:
L
KL
= KL(y||y
′
) (3)
where y
′
is the prediction of the teacher model
The total loss is computed as:
L = L
ce
+ λ
KL
L
KL
+ λ
distill
L
distill
(4)
with λ
KL
and λ
distill
are set to 0.5. This approach
enables DyVit to achieve efficiency by dynamically
pruning tokens while retaining the performance of a
full-token transformer.
4 IMPLEMENTATION DETAILS
The evaluation of the performances has been executed
through object-level reconstruction, using a subset of
the ShapeNet dataset (Chang et al., 2015) similar to
the Choi et al. (Choy et al., 2016) work. The sub-
set consists of 50,000 meshes along with their cor-
responding point clouds, covering 13 different object
Lightweight Transformer Occupancy Networks for 3D Virtual Object Reconstruction
411

classes. For each class 1/5 of the samples are used for
testing, 2/5 for validation, and the rest for training.
We used an NVIDIA GeForce GTX 1080 Ti to train
the model.
The hidden feature dimensions are 32 for the en-
coder and the decoder, the resolution of the x, y, and z
planes is set to 64, the batch size is 64 and the valida-
tion is done every 200 iterations. The learning rate of
the Adam optimizer is set at 10
−4
, and the betas val-
ues are set at β
1
= 0.9 and β
2
= 0.999. Moreover, we
set the depth of the DyVit to 2. We set the number of
heads of the transformer attention to 8 and the patch
size to 4. In this way, the model reaches convergence
after around 12k iterations.
The model employs a Sinusoidal Representation
Network (SIREN) as its activation function, intro-
duced by (Sitzmann et al., 2020). SIREN uses sine
functions as its basis, which naturally encode high-
frequency details, making it particularly effective for
processing inputs like images or 3D implicit represen-
tations. This periodic activation function ensures ef-
ficient training by carefully controlling the activation
values and spatial frequencies. The authors of SIREN
demonstrated its effectiveness across domains such as
images, videos, sound, and 3D data. Their method
outperforms traditional architectures, like those using
ReLU-based networks, in terms of accuracy, speed,
and the ability to capture fine details.
During inference, we apply Multiresolution Iso-
Surface Extraction (MISE) to extract meshes given
the occupancy values as inputs (Mescheder et al.,
2019). Because the objective of this work is to de-
crease the time of mesh generation, we are interested
in the inference time, which is expressed in seconds
per mesh.
Other metrics used for evaluation are essential to
determine the precision of the results and to under-
stand the consistency of the model. These metrics are
presented in the following:
Accuracy: calculated using the mean absolute
distance (MAD) between the dense points of the pre-
dicted mesh and the nearest surface of the ground
truth mesh. A lower MAD indicates better accuracy.
MAD
accuracy
=
∑
n
i
∥p
i
− g
i
∥
n
Where p
i
is the point of index i of the predicted mesh,
and g
i
is the point of index i of the ground truth.
Completeness: mean absolute distance between
the dense points of the ground truth and the surface of
the predicted mesh. The lower, the better.
MAD
completeness
=
∑
n
i
∥g
i
− p
i
∥
n
Volumetric Intersection over Union: the ratio
of the volume of the intersection to the volume of the
union of the ground truth and predicted mesh. The
higher, the better.
IoU =
Volumeo f Overlap
Volumeo f U nion
Chamfer-L1: the average of the accuracy and
completeness metrics. The lower, the better.
Cham f er-L1 =
MAD
accuracy
+ MAD
completeness
2
F-score: the average of precision and recall be-
tween the predicted mesh and the ground truth. The
higher, the better.
F-score = 2
precision · recall
precision + recall
5 EXPERIMENTS
In this section, we evaluate the performance of our
model compared to state-of-the-art approaches in
Section 5.1 and perform an ablation study in Section
5.2 to examine how changes in the network architec-
ture impact its results.
5.1 Comparison with Baseline Models
We compared our model with existing methods, re-
ferred to as baselines, on the ShapeNet dataset. As
shown in Table 1, our model generates a 3D mesh
in an average of 0.40 seconds per mesh, making it
the fastest among the models tested. As explained in
Section 4, for the scope of this work, whose purpose
is generating a roughly estimated mesh in less time
as possible (plausible mesh) to be refined in a subse-
quent stage, these metrics are less critical compared
to inference time. In Figure 3 we demonstrate the
consistency of our generated meshes by doing a vi-
sual comparison with the result of the DPConvONet
model.
Table 1: Performance comparison between our model and
state-of-the-art model. We can easily see that despite our
model giving non-optimal results, it is the fastest model
with respect to previous works.
Model Accuracy ↓ Completeness ↓ Chamfer-L
1
↓ F-score ↑ IoU ↑ Inference Time ↓
(Lionar et al., 2021a) 0.0047 0.0039 0.0043 0.9433 0.8879 0.86 s/mesh
(Lionar et al., 2021b) 0.0986 0.0999 0.0992 0.1851 0.7901 0.65 s/mesh
(Tang et al., 2021) 0.0097 0.0082 0.0090 0.7045 0.7429 0.45 s/mesh
(Tonti et al., 2024) 0.0058 0.0048 0.0053 0.9037 0.8499 0.48 s/mesh
Our 0.0342 0.0263 0.0302 0.3297 0.4924 0.40 s/mesh
Table 1 presents the evaluation metric values ob-
tained by the compared models. These values were
computed using the pre-trained weights provided in
the GitHub repositories of DPConvONet, Neuralblox,
SA-ConvONet, and LConvONet.
GRAPP 2025 - 20th International Conference on Computer Graphics Theory and Applications
412

Figure 3: Visual comparison of DPConvONet, our model, and the ground truth (shown from left to right). This comparison
demonstrates that our model produces consistent results, comparable to the meshes generated by the state-of-the-art method,
DPConvONet.
The obtained results present a boost of +53% (-
0.46s) with respect to DPConvONet, which is the
model with the best-resulting evaluation metrics, a
boost of +11% (-0.05s) for SA-ConvONet, which has
the lowest inference time, and a boost of +16% (-
0.08s) to LConvONet, which generates high-quality
details maintaining the inference time low. Fur-
thermore, when compared to Neuralblox, our model
demonstrates a significant improvement with a +38%
boost (-0.25s) and achieves higher evaluation metrics,
indicating more precise mesh generation.
5.2 Ablation Study
In Table 2 we propose different network configura-
tions. We replace the Siren activation function with
ReLU to analyze the performance difference between
the two architectures. We may notice that in this case,
the evaluation metrics values present a general wors-
ening, and the inference time increased. This con-
firms the results in (Sitzmann et al., 2020), which
demonstrate SIREN’s superior efficiency and accu-
racy compared to standard activation functions like
ReLU, Tanh, or GELU.
Table 2: Performance comparison between different config-
urations of the network.
Model Accuracy ↓ Completeness ↓ Chamfer-L
1
↓ F-score ↑ IoU ↑ Inference Time ↓
Our 0.0342 0.0263 0.0302 0.3297 0.4924 0.40 s/mesh
Network without Siren 0.0539 0.0312 0.0426 0.2293 0.3904 0.4521 s/mesh
Network with 3 DyViT 0.0361 0.026 0.031 0.3436 0.5124 0.4585 s/mesh
In the third row of the table, we present the results
given by the model when it processes separately the
3 planes with 3 different DyViTs. From such results,
we can notice that not only does the inference time in-
crease but processing the planes separately doesn’t re-
ally affect the accuracy of mesh generation. The mo-
tivation lies in the ability of the transformer to extract
both local and global features, which allows DyViT to
easily interpret the projections onto the 3 planes also
when such planes are overlapped.
These findings highlight that the use of SIREN
and overlapping projection planes contributes signifi-
cantly to the efficiency and quality of our model.
6 CONCLUSIONS
This study explores the integration of an efficient vi-
sion transformer, such as DyViT, into a Convolutional
Occupancy Network, replacing three U-Nets used for
processing point cloud projections. The transformer’s
ability to handle both local and global features allows
for overlapping projection planes, significantly allevi-
ating bottlenecks between the encoder and decoder.
We also highlight the importance of the SIREN
activation function in preserving a high level of detail
within the model’s output.
Looking ahead, we aim to further reduce inference
time by refining the mesh representation process. In
computer graphics, 3D artists often create a coarse ob-
ject shape and enhance it with bump, normal, and par-
allax mapping to maintain detail. Inspired by this, our
Lightweight Transformer Occupancy Networks for 3D Virtual Object Reconstruction
413

future objective is to represent the point cloud using a
single projection plane and use our model to densify
it. This approach would enable us to generate paral-
lax mapping for a coarse 3D mesh, maintaining visual
detail while optimizing computational efficiency.
REFERENCES
Alhaija, H. A., Mustikovela, S. K., Mescheder, L., Geiger,
A., and Rother, C. (2017). Augmented reality meets
deep learning for car instance segmentation in urban
scenes. In British machine vision conference, vol-
ume 1.
Chabra, R., Lenssen, J. E., Ilg, E., Schmidt, T., Straub,
J., Lovegrove, S., and Newcombe, R. (2020). Deep
local shapes: Learning local sdf priors for detailed
3d reconstruction. In Computer Vision–ECCV 2020:
16th European Conference, Glasgow, UK, August 23–
28, 2020, Proceedings, Part XXIX 16, pages 608–625.
Springer.
Chang, A. X., Funkhouser, T., Guibas, L., Hanrahan,
P., Huang, Q., Li, Z., Savarese, S., Savva, M.,
Song, S., Su, H., et al. (2015). Shapenet: An
information-rich 3d model repository. arXiv preprint
arXiv:1512.03012.
Chibane, J., Alldieck, T., and Pons-Moll, G. (2020). Im-
plicit functions in feature space for 3d shape recon-
struction and completion. In Proceedings of the
IEEE/CVF conference on computer vision and pattern
recognition, pages 6970–6981.
Choy, C. B., Xu, D., Gwak, J., Chen, K., and Savarese,
S. (2016). 3d-r2n2: A unified approach for single
and multi-view 3d object reconstruction. In Computer
Vision–ECCV 2016: 14th European Conference, Am-
sterdam, The Netherlands, October 11-14, 2016, Pro-
ceedings, Part VIII 14, pages 628–644. Springer.
Curless, B. and Levoy, M. (1996). A volumetric method for
building complex models from range images. In Pro-
ceedings of the 23rd annual conference on Computer
graphics and interactive techniques, pages 303–312.
Genova, K., Cole, F., Sud, A., Sarna, A., and Funkhouser,
T. (2020). Local deep implicit functions for 3d shape.
In Proceedings of the IEEE/CVF Conference on Com-
puter Vision and Pattern Recognition, pages 4857–
4866.
Lionar, S., Emtsev, D., Svilarkovic, D., and Peng, S.
(2021a). Dynamic plane convolutional occupancy net-
works. In Proceedings of the IEEE/CVF Winter Con-
ference on Applications of Computer Vision, pages
1829–1838.
Lionar, S., Schmid, L., Cadena, C., Siegwart, R., and Cra-
mariuc, A. (2021b). Neuralblox: Real-time neural
representation fusion for robust volumetric mapping.
In 2021 International Conference on 3D Vision (3DV),
pages 1279–1289. IEEE.
Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S.,
and Geiger, A. (2019). Occupancy networks: Learn-
ing 3d reconstruction in function space. In Proceed-
ings of the IEEE/CVF conference on computer vision
and pattern recognition, pages 4460–4470.
Papa, L., Russo, P., and Amerini, I. (2023). Meter: a mobile
vision transformer architecture for monocular depth
estimation. IEEE Transactions on Circuits and Sys-
tems for Video Technology, pages 1–1.
Park, J. J., Florence, P., Straub, J., Newcombe, R., and
Lovegrove, S. (2019). Deepsdf: Learning continu-
ous signed distance functions for shape representa-
tion. In Proceedings of the IEEE/CVF conference on
computer vision and pattern recognition, pages 165–
174.
Patil, A. G., Qian, Y., Yang, S., Jackson, B., Bennett, E.,
and Zhang, H. (2023). Rosi: Recovering 3d shape
interiors from few articulation images. arXiv preprint
arXiv:2304.06342.
Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M.,
and Geiger, A. (2020). Convolutional occupancy net-
works. In Computer Vision–ECCV 2020: 16th Euro-
pean Conference, Glasgow, UK, August 23–28, 2020,
Proceedings, Part III 16, pages 523–540. Springer.
Qi, C. R., Su, H., Mo, K., and Guibas, L. J. (2017). Point-
net: Deep learning on point sets for 3d classification
and segmentation. In Proceedings of the IEEE con-
ference on computer vision and pattern recognition,
pages 652–660.
Rao, Y., Liu, Z., Zhao, W., Zhou, J., and Lu, J. (2023). Dy-
namic spatial sparsification for efficient vision trans-
formers and convolutional neural networks. IEEE
Transactions on Pattern Analysis and Machine Intel-
ligence, 45(9):10883–10897.
Sitzmann, V., Martel, J. N. P., Bergman, A. W., Lindell,
D. B., and Wetzstein, G. (2020). Implicit neural rep-
resentations with periodic activation functions.
Tang, J., Lei, J., Xu, D., Ma, F., Jia, K., and Zhang, L.
(2021). Sa-convonet: Sign-agnostic optimization of
convolutional occupancy networks. In Proceedings of
the IEEE/CVF International Conference on Computer
Vision, pages 6504–6513.
Tonti, C. M., Papa, L., and Amerini, I. (2024). Lightweight
3-D Convolutional Occupancy Networks for Virtual
Object Reconstruction . IEEE Computer Graphics
and Applications, 44(02):23–36.
Touvron, H., Cord, M., El-Nouby, A., Verbeek, J., and
J
´
egou, H. (2022). Three things everyone should know
about vision transformers. In European Conference
on Computer Vision, pages 497–515. Springer.
Urmson, C., Anhalt, J., Bagnell, D., Baker, C., Bittner, R.,
Clark, M., Dolan, J., Duggins, D., Galatali, T., Geyer,
C., et al. (2008). Autonomous driving in urban envi-
ronments: Boss and the urban challenge. Journal of
field Robotics, 25(8):425–466.
Virtanen, J.-P., Daniel, S., Turppa, T., Zhu, L., Julin, A.,
Hyypp
¨
a, H., and Hyypp
¨
a, J. (2020). Interactive dense
point clouds in a game engine. ISPRS Journal of Pho-
togrammetry and Remote Sensing, 163:375–389.
Zhu, Y., Mottaghi, R., Kolve, E., Lim, J. J., Gupta, A., Fei-
Fei, L., and Farhadi, A. (2017). Target-driven visual
navigation in indoor scenes using deep reinforcement
learning. In 2017 IEEE international conference on
robotics and automation (ICRA), pages 3357–3364.
IEEE.
GRAPP 2025 - 20th International Conference on Computer Graphics Theory and Applications
414