bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024.

The copyright holder for this preprint

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

A Toolbox for Generating Multidimensional 3-D Objects with Fine-

Controlled Feature Space: Quaddle 2.0
Xuan Wen1,2, Leo Malchin1, Thilo Womelsdorf1,2,3
1
Department of Psychology, Vanderbilt University, Nashville, TN 37240
2
Vanderbilt Brain Institute, Nashville, TN 37240
3
Department of Biomedical Engineering, Vanderbilt University, Nashville, TN 37240

Keywords: Object recognition • Virtual Reality • Augmented Reality • Video Games

• Naturalistic Stimuli • Psychophysical Experiments

Corresponding Authors: Xuan Wen ([email protected]) and Thilo Womelsdorf

([email protected])

Abstract
Multidimensional 3D-rendered objects are an important component of vision research and video-
gaming applications, but it has remained challenging to parametrically control and efficiently
generate those objects. Here, we describe a toolbox for controlling and efficiently generating 3D
rendered objects composed of ten separate visual feature dimensions that can be fine-adjusted
using python scripts. The toolbox defines objects as multi-dimensional feature vectors with
primary dimensions (object body related features), secondary dimensions (head related features)
and accessory dimensions (including arms, ears, or beaks). The toolkit interfaces with the freely
available Blender software to create objects. The toolbox allows to gradually morph features of
multiple feature dimensions, determine the desired feature similarity among objects, and
automatize the generation of multiple objects in 3D object and 2D image formats. We document
the use of multidimensional objects in a sequence learning task that embeds objects in a 3D-
rendered augmented reality environment controlled by the gaming engine unity. Taken together,
the toolbox enables the efficient generation of multidimensional objects with fine control of low-
level features and higher-level object similarity useful for visual cognitive research and immersive
visual environments.
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Introduction

Investigating cognition requires tools that can precisely manipulate and control the features of
stimuli presented to participants. Multidimensional, feature-controlled object sets have become
indispensable in cognitive research for probing perception, recognition, and categorization
processes (Arnott et al., 2008; Biederman & Gerhardstein, 1993; Bowman & Zeithamova, 2020;
Chuang et al., 2012; Ghazizadeh et al., 2016; Knutson et al., 2012; Mercer & Duffy, 2015; Pusch
et al., 2023; Todd, 2004; Wallraven et al., 2014; Wong et al., 2009). In previous research, Watson
et al., 2019 introduced a toolkit that generated 3D-rendered objects with fine user control of up to
five different visual feature dimensions (Watson et al., 2019). Multidimensional objects generated
with this toolkit, so called Quaddle objects, have been effectively utilized to explore various
aspects of cognitive processing, such as object recognition, discrimination, visual search,
attentional set shifting and feature learning (Banaie Boroujeni et al., 2022; Boroujeni et al., 2022;
Hassani et al., 2023; Kemp et al., 2024; Womelsdorf, Thomas, et al., 2021; Yazdanpanah et al.,
2024). Moreover, 3D-rendered multidimensional objects are used in cognitive research using
augmented reality (AR) and virtual reality (VR) that provide immersive environments with the
aim to mimic real-world experiences (Smith, 2019). For these dynamic and sometimes interactive
settings 3D rendering of objects can enhance task performance and have been used in contexts
studying object identification, spatial manipulation and navigation, object memory recall and
learning (Corriveau Lecavalier et al., 2020; Howett et al., 2019; McIntire et al., 2014).

Despite the usefulness of 3D rendered objects in the described settings, researchers face challenges
in generating well controlled 3D objects due to technical constraints and the complexity associated
with 3D modeling software. Addressing this situation, we introduce the Quaddle 2.0 toolbox, an
evolution of the original Quaddle toolkit designed to enhance utility and flexibility of object
generation for dynamic visual and cognitive science studies and video-game contexts (Watson et
al., 2019). Quaddle 2.0 expands upon its predecessor by increasing the available feature space
(from five to ten visual feature dimensions) (Figure 1A), while still allowing generating simpler
Quaddle 1.0 objects that has less feature dimensions (Figure 1B). In addition, the Quaddle 2.0
toolbox newly establishes the use of Blender - a free, open-source 3D creation suite compatible
with macOS and Windows – to interface flexibly with python scripts, allowing an efficient
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

automatized control over low-level object features (gradual feature value changes) and higher-
level 3D object characteristics (e.g. light viewing angle).

Figure 1. Examples of Quaddle objects. (A) Quaddle 2.0 objects varying in ten different feature
dimensions. (B) Simpler objects that vary in features of four dimensions (body shape, arm angle, body
colors, body pattern) shows the toolbox allows reproducing Quaddle 1.0 objects that were generated
with a different software rendering suite (Studio DS Max). (B) Quaddle 2.0 objects that are more similar
to each other by setting the similarity score > 8.

Here, we (i) detail the Quaddle 2.0 framework, (ii) introduce the scripting pipeline for generating
large object sets, (iii) illustrate how users can control the feature similarity among objects, (iv)
demonstrate the gradual morphing of multiple object features, and (v) showcase the use of Quaddle
2.0 objects in an experimental environment programmed with the unity game engine to evaluate
how non-human primates learn the sequential ordering of visual objects. Example code (in python
and MATLAB) and a documentation with detailed instructions for using and customizing Quaddle
2.0 objects is freely available online (see Appendix).

Method
Composition of Quaddle 2.0 objects. Quaddle 2.0 objects vary in features of ten pre-configured
object dimensions (Figure 2). The default object contains five main parts: head, body, arm, ear,
and beak. The head and body can be covered with surface color and patterns, and the body can
have pre-configured, colorful fractal images projected on its center. The default arm, ear and beak
are colored white and can be adjusted to have different length, shape and bending angle. Arm, ear
and beak are considered minor feature dimensions that can be omitted from the objects (Figure 2).
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Figure 2. Object dimensions and example feature values pre-configured in the Quaddle 2.0
toolbox. Feature dimensions are categorized into primary dimensions (body-related), secondary
dimensions (head-related), and minor dimensions (accessories), organized by size and front-view
visibility. The ‘weight’ row denotes how a feature dimension is weighted when calculating the similarity
of objects (see similarity score calculations).

Generating 3D objects using lists of pre-configured object features. Quaddle 2.0 objects are
rendered using Blender (blender.org), a free and open-source 3D modeling software suite. Objects
generated for figures in this article used Blender software version 3.4.0. Generating Quaddle 2.0
objects follows a structured pipeline implemented through a series of Python scripts within a
Blender project file as outlined in Figure 3. All Python scripts are freely available online (see
Appendix). The process begins with user-specified information about the desired feature
dimensions and feature valuer per dimension, which are provided via a command-line tool or in a
text file. The input is parsed using a parser function, which extracts the values for each visual
feature of the different object dimensions. The core of the generation process involves several
scripts that sequentially build the Quaddle as surveyed in Figure 2. Each component of the
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Quaddle is generated through a series of modular functions. The process begins with the creation
of the main body, followed by the attachment of the head. Subsequently, additional functions
append the ears to the head, add arms to the body, and attach a beak to the head. This modular
approach allows for precise control over each component of the Quaddle, enabling systematic
manipulation of individual features while maintaining overall structural consistency. The body of
the Quaddle object is initially generated from the default mesh shapes in Blender, including sphere,
cube, cone, and cylinder, and then molded into the user-specified desired body shape. The arms,
beak and ears are initially generated as straight cylinders and molded into the desired angle, shape,
and length, and which allows gradual morphing of object features by specifying intermediary
values in bended angle, size of shape and length. In the feature set pre-configured in the Quaddle
2.0 toolbox, each Quaddle has at most two arms, two ears and one beak (Figure 4). There are three
angles for arm (bend up, bend down, straight), three shape for ear (straight, pointy and blunt), and
three lengths for ear (regular, long, short). Different sub-features can be combined freely.
Additional feature dimensions and feature values can be realized by modifying the pre-configured
python code.

Figure 3. Generation pipeline for Quaddle 2.0 objects. The automated pipeline for generating
Quaddle models using the makeQuaddle.blend Blender project file, which contains predefined Python
functions. User-specified input (either via a command-line tool or a text file) is parsed
by tableParser.py, which extracts the parameters used to generate the object with the
script GenerateObject.py. The main pipeline splits into two key branches: (1) 2D Textures,
where functions apply body and head colors, patterns, and body fractals; and (2) 3D Models, where
scripts build the respective morphological elements. Helper functions handle lighting, object
positioning, and naming before exporting to the user-specified destination.
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Generating 2-D textures and surface features. Following the structural assembly, 2D textures
are applied to the 3D model. This script addImageTextures.py manages the application of body
color, pattern, and fractal designs, as well as head color and pattern. The surface colors and patterns
are imported from PNG files. The default gray color is the same for all objects, whereas the other
colors are chosen within the CIE L*c* h* space such that the L* and c* values (luminance and
saturation, respectively) are held constant, but h* values (hue) vary by 30°, meaning that there is
a small difference in hue between the two colors, but not in the other components of the colors.
The texturing process employs a UV project modifier, a specialized tool in Blender that projects
2D images onto the 3D object's surface. This modifier projects images from the front of the object,
minimizing distortion when viewed frontally and ensuring symmetrical pattern application on both
front and back. Two separate UV project modifiers are utilized for background patterns and surface
fractals separately, each with a distinct projector object for optimal surface coverage. The process
involves creating a material with a node-based shader network. Two image textures are
incorporated and combined using a mixing operation, allowing for complex surface patterns. The
first texture applies directly to the object surface, while the second serves as both a secondary color
source and a blending mask. This approach prevents distortions of frontal views, facilitates
symmetrical patterns, and allows control over texture placement and scale. By manipulating UV
maps and projector properties, the process achieves various texture effects, from uniform coverage
to complex, position-dependent patterns. The projection method can be similarly applied across
different object geometries, which allows using the same complex visual pattern across a diverse
set of object shapes.

Figure 4. Illustration of gradual morphological changes between feature dimensions using the
Quaddle 2.0 toolbox. The objects transition in body shape (from conical to cylindrical to rectangular),
arm angles and orientations, color and surface pattern (from pink and purple horizontal stripes to green
and blue vertical stripes), and beak shape (blunt to pointy).
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

The Quaddle 2.0 toolbox allows projecting fractal images onto the object body or to render various
body surface patterns. Fractal images are used because they provide a large and abstract image
space that is unfamiliar to experimental subject. The fractals were generated following previously
published methods (Ghazizadeh et al., 2016) using point-symmetrical polygons, where parameters
such as size, edges, and color were randomly chosen to create distinct visual shapes. This method
ensures that a large number of distinguishable fractals are available.

With regard to patterns for the body surface of objects, the Quaddle 2.0 toolbox has five pre-
configured patterns available: solid (default), horizontal stripes, diagonal stripes, vertical stripes,
and grid. Each pattern is generated as a 1200x1200 PNG file, with stripe widths set to 150 pixels
to ensure the patterns and their associated colors remain clear and distinguishable in the final
stimuli. While users can select any combination of colors for the patterns from the prepared asset
folder, the current code defaults to combining a neutral grey with a user-specified secondary color
(for examples, see Figure 1). This design ensures that the colors are well visible and have apparent
contrast in the patterned image.

Refinement and exporting 3D rendered objects. Post-processing steps are then executed to
refine the generated Quaddle objects. These steps include creating homogeneous lighting
conditions and exporting all modified textures, ensuring consistent visual quality across stimuli.
The process also involves locating the center of the final object, repositioning and resizing it as
necessary to maintain standardized dimensions across all generated Quaddle objects. Additionally,
a unique identifier is assigned to each generated object, facilitating organized data management in
subsequent experimental procedures.

The final stage constructs the complete object based on all specified parameters and outputting it
in the required format to a user-defined destination. The Blender Python script that can read
multiple object definitions at once and create large sets of object, exporting them either as image
files (e.g., PNG), or 3D shape files (e.g., FBX, GLTF) with any specified distance, height, and
rotation. The default output centers the object at the middle points of height and width of each
object. The camera distance is adjusted based on the size of the objects with the aim to maintain a
consistent apparent size across objects.
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Controlling the similarity of multidimensional objects. Quaddle 2.0 objects can share variable
numbers of features, which we exploit by calculating a similarity score between objects. Users can
set a similarity score in the make_quaddle.m MATLAB script when generating two or more
objects. Similarity is calculated by comparing the values of each feature between two objects, with
different weights assigned to each dimension. Visually more prominent object dimensions weight
strongest (weights of body shape: 2; body color: 1.5, body pattern, 1.5), while less prominent
secondary dimensions weight less (head shape, head color and head pattern: weight 0.75), and
minor dimensions weight least (Arm, Fractal, Ears, Beak: weight 0.25). Using these feature
dimension weights the similarity score consider
pairs of Quaddle objects 𝑸! and 𝑸" , defined by their dimension 𝑸! = [𝑞!! , 𝑞!" , … , 𝑞!# ] and 𝑸" =
[𝑞"! , 𝑞"" , … , 𝑞"# ]. A binary vector indicates whether the values in each dimension are the same is
defined as 𝑩 = [𝑏! , 𝑏" , … , 𝑏# ] where 𝑏$ = 1 if 𝑞!$ = 𝑞"$ , else 𝑏$ = 0. Weights are predefined for
each feature dimension 𝑖 : 𝑾 = [𝑤! , 𝑤" , … , 𝑤# ]. The Similarity Score S is calculated as 𝑆 =
(𝑾)% (𝑩) or equivalently as 𝑆 = ∑#$&! 𝑤$ 𝑏$ . Using the pre-configured Quaddle 2.0 objects, the
similarity score ranges from 0 (most dissimilar) to 11.75 (most similar). We validate the similarity
score by comparing the similarity of object pairs as defined by the score with other quantitative

Figure 5. Comparison of object similarity scores A: Scatter plot showing the relationship between
normalized SSIM values and similarity scores, including the distribution of both scores. The regression
line is y=0.206x+0.266 with 𝑟 ! =0.047. B: Scatter plot illustrating the correlation between normalized
EfficientNetB0 similarity values and similarity scores. The regression line is y=0.301x+0.332
with 𝑟 ! =0.093. C: Scatter plot depicting the association between normalized VGG16 similarity values
and similarity scores. The regression line is y=0.337x+0.255 with 𝑟 ! =0.127. Each figure is based on a
random draw of 1000 pairs of 300 randomly generated Quaddle images.
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

similarity estimates (see below and Figure 5) and by empirically evaluating the perceptual
confusability of objects with high versus low similarity scores (see below and Figure 6).

Experimental paradigm for serial ordering of Quaddle objects with controlled feature
similarity. To demonstrate a use case for the 3D rendered Quaddle 2.0 objects we generated
multiple sets of Quaddle objects and trained four rhesus macaque monkeys to learn pre-defined
temporal (serial) orders of five objects in each those object sets in a sequence learning task. We
defined the sixth object as a distractor with a high similarity to only one of the other objects to

Figure 6. Experimental paradigm and results. A: Illustration of Quaddle 2.0 objects that are
perceptually dissimilar (left) and similar (right) as quantified by the similarity score (see text for
details). B: Each trial presented six objects. Monkeys learned to touch five objects in a pre-determined
order A-B-C-D-E and avoid a distractor object. A correct choice led to visual feedback (yellow halo)
and incremented the slider position on top of the screen. Subjects had 10 maximum number of errors
in exploring one sequence. Successfully completing one sequence would lead to fluid rewards. In each
trial, distractor object has higher similarity to either 2ed or 4th object in the sequence, compared to lower
similarity with other objects. C: Learning curves for completing sequences with ≤ 10 errors (y-axis).
Symbols show avg. trial and SE to reach 80% completion for each subject and overall. D: Proportion
of distractor choices across trials. The mean value is shown as the rightmost data point (Mean ± 95%
CI: 0.090 ± 0.010). E: Distractor choices at each ordinal position when distractor is similar to object at
the 2nd ordinal position and when distractor is similar to object in 4th ordinal position. Two-proportion
Z-tests were applied at each ordinal position for comparing the difference. Stars denote sign. level.
(Ordinal Position: Before & After Swap, Mean ± 95%CI) 1: 0.13 ± 0.01; 1: 0.08 ± 0.01; 2: 0.29 ± 0.01;
2: 0.20 ± 0.01; 3: 0.19 ± 0.01; 3: 0.22 ± 0.01; 4: 0.22 ± 0.01; 4: 0.34 ± 0.01; 5: 0.17 ± 0.01; 5: 0.16 ±
0.01. Figure panels B and C were adapted from Wen et al., 2024.
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

illustrate the distractor is confused specifically with an object that has a high similarity to it as
defined by the similarity score. All animal and experimental procedures used in this study
complied with the National Institutes of Health Guide for the Care and Use of Laboratory Animals
and the Society for Neuroscience Guidelines and Policies and were approved by the Vanderbilt
University Institutional Animal Care and Use Committee.

The details of the experimental set-up and the timing of task events are described in detail in (Wen
et al., 2024). In brief, four adult male rhesus monkeys performed the sequence learning task in
their housing cage using cage-mounted touch screen Kiosk stations (Womelsdorf, Thomas, et al.,
2021). Visual display, behavioral response registration, and reward delivery was controlled by the
Multi-Task Suite for Experiments (M-USE) (Watson et al., 2023). M-USE is a free software
platform with pre-configured tasks (https://m-use.psy.vanderbilt.edu) that displays 3D rendered
objects using the Unity3D program (here: Unity version 2020.3.40F1). M-USE displayed the
objects on an Elo 2094L 19.5" LCD touchscreen with a refresh rate of 60 Hz and a resolution of
1,920 × 1,080 pixels, rendered at approximately 3 cm on the screen. Experimental sessions tested
at least eight unique sequences, each with a unique set of six Quaddle 2.0 objects. Sets of objects
were generated by randomly assigning objects different features from up to ten different feature
dimensions (see Figure 1). Five of the objects were defined to be dissimilar to make them easy to
distinguish, while a sixth object was a distractor that was not part of a sequence but shared features
with either the object at the second or the fourth serial order (Figure 6A). The similarity of the
distractor with selected objects of the sequence was used to validate that the similarity of Quaddle
2.0 objects translated into perceptual confusability among objects.

As illustrated in Figure 6B, the task presented on each trial six objects equidistant to each other at
random locations on a circle equidistant from the center of the screen. Five of the objects were
assigned a unique ordinal temporal position in the sequence of objects A-B-C-D-E, while a sixth
object was a sequence-irrelevant distractor. In each trial subjects chose objects by touching them
and received either positive feedback (a yellow halo and high pitch sound) for correct, or negative
feedback (a transient grey halo and low pitch sound) for choosing an object at an incorrect temporal
position. After an erroneous choice, subjects had to re-choose the last correctly chosen object
before making a choice of an object believed to be the next object in the sequence. Each trial
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

allowed a maximum of ten errors to complete the sequence and receive fluid reward. If the
sequence was not completed the objects were removed from the screen and a new trial was started
with objects rearranged to new random locations. Subjects performed each sequence for a
maximum of fifteen trials to allow for the gradual acquisition of the sequential ordering and
additional trials to exploit the learned sequence. For each correctly chosen object, the position of
the slider of a progress bar at the top of the screen stepped forward. Successful completion of a
sequence always completed the slider progress bar and resulted in a fluid reward. When a sequence
was not completed no fluid reward was given and the slider progress bar was reset for the
subsequent trial with the same objects at new random locations. Choices of the distractor were
registered as incorrect and not rewarded and required re-touching the last correctly chosen object.

Hardware for automatic generation of Quaddle 2.0 objects. The Quaddle 2.0 toolbox has been
developed and tested on an Apple MacOS system, but with minor adjustments it will be compatible
with other operating systems that run Python and the Blender software (the tested Blender software
version was 3.4.0). For the experiment, Quaddle 2.0 objects were generated with an Apple
MacBook Pro 2023 with Apple M2 Max Chip and 32 GB Memory using MATLAB scripts
(MATLAB 2022b) and python version 3.10.8. The objects presented here were generated with
python scripts that connect with Blender version 3.4.0 (see Figure 3). Links to a comprehensive
documentation and to the resources reproducing the objects are described in the Appendix.

Results

Automatized generation of multidimensional objects. The Quaddle 2.0 toolbox connects the
object creation software Blender with MATLAB via a command-line interface, enabling users to
generate objects bypassing Blender’s interface. This feature enables the automatized generation of
large object databases with thousands of pre-configured objects. The MATLAB script
make_quaddle.m (see Appendix) allows users to specify the number of objects, set the object
similarity, and determine which features to keep or vary, providing high flexibility and utility for
any application that requires multiple new objects across multiple cognitive experiments or
augmented reality worlds. The script will generate an object table file for each specified object. A
standard laptop is sufficient to generate objects. Objects displayed in Figure 1, 2, 4, and 6 were
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

generated on an Apple MacBook Pro (2023) equipped with the Apple M2 Max chip and 32 GB of
memory.

Computing efficiency for different formats. The pre-configured scripts (see Appendix) allow
users to select any combination of 3D object formats (FBX, GLTF) or a 2D image output (PNG).
The time taken to generate each object varies depending on the selected output format. When
generating GLTF files only, the average time to generate and save an object was 1.32 seconds per
object. For FBX files only, the time increased slightly, averaging 1.55 seconds per object.
Generating PNG images took longer, with an average time of 2.35 seconds per object. When
generating GLTF, FBX, and PNG files simultaneously, the time increased to 2.42 seconds per
object, primarily due to the additional time required to render the 2D image compared to saving
3D object formats directly.

Improvements of the Quaddle 2.0 platform over previous versions. Quaddle 2.0 objects
provide more feature dimensions and a larger feature space than Quaddle 1.0 objects (Watson et
al., 2019), but can be used with less than ten feature dimensions, which allows reproducing objects
with four feature dimensions similar to Quaddle 1.0 objects as shown in Figure 1C. The Quaddle
1.0 framework had other limitation that the new framework addresses. Quaddle 1.0 objects were
rendered and generated from within the Studio DS Max software on windows PCs. By using the
Blender software on PC or MacOS operating systems, and interfacing python scripts to Blender
the Quaddle 2.0 platform moves beyond these limitations. This interface enables rapid ‘on-demand’
generation of objects for applications that require multiple new objects with random combinations
of features as is typical for multiple types of learning studies that introduces novel objects over
hundreds of experimental sessions (Banaie Boroujeni et al., 2022; Treuting et al., 2024;
Womelsdorf, Watson, et al., 2021).

Gradual morphing between features of multiple dimensions. Category learning studies have
demonstrated that fine, gradual manipulations of object features is needed for controlling the
discriminability of relevant dimensions in cognitive tasks (Apostel & Rose, 2021; Folstein et al.,
2012; Minda & Smith, 2001). The Quaddle 2.0 platform enables such fine control of low level
features and provides morphing functionality that can be applied to multiple features
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

simultaneously as illustrated in Figure 4. Gradual changes of feature values include

morphing body shapes, adjusting the angle of stripe patterns, and modifying color shading.
Additionally, Quaddle 2.0 offers precise control over arm angle and the bluntness or sharpness of
arm tips (Figure 4). Code and instructions for morphing specific color and pattern are available
on the website (see Appendix).

Quantitative evaluation of the similarity score of objects. The MATLAB object generation
script make_quaddle.m (see Appendix) accompanying the Quaddle 2.0 toolbox calculates the
similarity among objects by determining how many features are on average shared among objects
in a given object set, weighting differently shared primary, secondary, and accessory features (see
Methods). We quantitatively evaluated this similarity measure by comparing it with two existing
alternative similarity measures. First, we calculated the Structural Similarity Index (SSIM), which
measures the visual similarity between images by comparing luminance, contrast, and structure at
the pixel level (Wang et al., 2004). We calculated the SSIM of Quaddle 2.0 objects exported as
PNG images by applying the python package skimage. We computed the SSIM and the similarity
score for 1000 pairs of object images randomly drawn from 300 randomly generated Quaddle 2.0
objects. Both, the SSIM and the similarity score were moderately positively correlated with similar
average SSIM and similarity scores (SSIM: mean:0.4, variance: 0.03; Similarity Score: mean: 0.35,
variance: 0.03) (Figure 5A). Secondly, we quantified how gradually increasing similarity scores
of Quaddle 2.0 objects relate to a similarity score estimated by pretrained deep neural network
models. These models process input data through multiple layers to extract meaningful patterns
and features. At one or more hidden layers, the input data is transformed into dense vector
representations known as feature embeddings. These embeddings are numerical outputs that
capture the essential characteristics of images in a compact, multidimensional space (LeCun et al.,
2015). We used the TensorFlow library (Abadi et al., 2016) to extract the feature vectors from the
VGG16 model and the EfficientNetB0 model, which are pre-trained on large datasets like
ImageNet (Simonyan & Zisserman, 2014; Tan & Le, 2019). Feature embedding vectors of two
images were extracted and cosine similarity between two vectors were calculated as similarity
score. The networks allow users to choose similarity metrics based on low-level structural
comparisons or evaluation of high-level features-similarity. We found that the similarity score
shows significant, albeit moderately positive correlations with the EfficientNetB0 similarity value
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

with an 𝑟 " value of 0.09 (slope = 0.301; intercept = 0.332) (Figure 5B). Similarly, the VGG16
similarity value was also moderately correlated, with an 𝑟 " value of 0.127 (slope = 0.337; intercept
= 0.255) (Figure 5C). These results show that scoring the similarity of Quaddle 2.0 objects based
on shared feature spaces with a differential weighting of primary, secondary and minor feature
dimensions is weakly but positively correlated with quantitative assessments of low-level feature
similarity metrics.

Empirical evaluation of Quaddle 2.0 objects and their similarity score. To test how subjects
learn to associate different Quaddle 2.0 objects that are dissimilar and to test how higher similarity
scores correspond to perceptual confusability of objects we tested rhesus monkeys on a sequence
learning task. The testing was part of a larger empirical study (Wen et al., 2024) and included 125
testing sessions across four adult male rhesus monkeys (Subject B: 13, Subject J: 40, Subject K:
59, Subject S: 13). Each session contained on average 24 (Range: 9-26) sequence learning blocks.
For each of the learning blocks on each of the sessions, six new Quaddle 2.0 objects were generated
(five objects that were part of the sequence, and one distractor object). Subjects learned the
predetermined 5-object sequences. They reached criterion performance (completing a sequence
with maximally 15 choices, i.e. ten or less erroneous choices) within 5.34 ± 0.37 trials (Subject B:
9.55 ± 1.62; J: 6.88 ± 0.68; K: 3.61 ± 0.34; S: 5.18 ± 0.93) (Figure 6C). Learning was accompanied
by a steady decrease of the erroneous choices of the distractor (Figure 6D). In the experiment, the
distractor object resembled either the second or fourth object of the sequence, and distractor errors
occurred significantly more often at positions at which the sequence-relevant object and the
distractor was similar, i.e. the distractor object was more easily confused with objects that had
higher similarity scores (Figure 6E).

Discussion
We introduced the Quaddle 2.0 toolbox for generating 3D rendered objects with multiple
controllable feature dimensions for cognitive research and video gaming applications. Building on
the foundation of the original Quaddle 1.0 framework, Quaddle 2.0 enhances stimulus generation
by expanding number of independently varying feature dimensions and improving the script-based
accessibility through the use of Blender, an open-source 3D modeling platform with Python
support. Quaddle 2.0 allows using simple configuration text files to define objects and fine adjust
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

low-level object features, which makes it easier to create custom object sets and automatize the
generation of large object sets. Quaddle 2.0 supports various 3D model formats which enable their
use in augmented and virtual reality platforms. Together, these toolbox features facilitate dynamic
and immersive experimental designs.

It can replicate the key features from the original Quaddle framework (Watson et al., 2019), which
has been successfully tested in experiments with both humans and non-human primates (Banaie
Boroujeni et al., 2022; Boroujeni et al., 2022; Hassani et al., 2023; Kemp et al., 2024; Womelsdorf,
Thomas, et al., 2021; Yazdanpanah et al., 2024). Extending this earlier framework, we have shown
how the Quaddle 2.0 were used in sequence learning that required hundreds of new, unique objects
in each experimental session, and which controlled the similarity among objects. Non-human
primates learned to discriminate and serially order the objects and were confused by distractors
when it shared a high feature similarity with target objects (Figure 6). Taken together, this example
use case suggests that Quaddle 2.0 is a versatile, scalable, and user-friendly tool that enables
cognitive research in 3D rendered environments.
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Appendix

• Python scripts and install instruction of the pipeline generating Quaddle 2.0 objects (shown in
Figure 3) are available at https://github.com/xwen1765/blender-quaddle. Setting up the toolbox
involves three steps: 1) Download Blender 3.4.0 and python code from the GitHub repository on
the local computer. 2) Open Blender software, change preference as in the instruction. 3) Open the
file makeQuaddle.blend and run the script parser.py, an example Quaddle will be created and can
be viewed in Blender. To automatically generate object tables and generate 3D object sets with
user defined similarity between objects, the user can use MATLAB. The GitHub repository
includes the files make_quaddle.m and test_example.m that showcases how the configuration
variables are set, object table files are generated and the parser.py is called to render and generate
the Quaddle 2.0 objects.
• Website documentation about MATLAB scripts is available in the text file “Documentation-
MATLAB-make_quaddle.pdf” on the website https://m-use.psy.vanderbilt.edu/quaddles-2/.
• Website installation guide is available in the
• An example of the morphing functionality using specific color and pattern is available on the
website https://github.com/xwen1765/blender-quaddle.
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Author Contributions

XW and TW were responsible for conceptualizing the research; XW developed the toolbox and
designed the object features; XW and LM contributed programming code and generated the results
and figures; TW supervised the study. All authors contributed to and have approved of the final
version of the manuscript.

Funding
This work was supported by the National Institute of Mental Health (R01MH123687). The funders
had no role in study design, data collection and analysis, the decision to publish, or the preparation
of this manuscript.
Data availability Not applicable.
Code availability All MATLAB and python scripts of the Quaddle 2.0 toolbox are publicly
available via https://github.com/xwen1765/blender-quaddle
Declarations
Conflict of Interest Statement The authors declare no competing financial interests.
Ethics approval Not applicable.
Consent to participate Not applicable.
Consent for publication Not applicable.
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

References
Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G. S., Davis, A.,

Dean, J., Devin, M., Ghemawat, S., Goodfellow, I., Harp, A., Irving, G., Isard, M., Jia, Y.,

Jozefowicz, R., Kaiser, L., Kudlur, M., … Zheng, X. (2016). TensorFlow: Large-Scale

Machine Learning on Heterogeneous Distributed Systems (Version 2). arXiv.

https://doi.org/10.48550/ARXIV.1603.04467

Apostel, A., & Rose, J. (2021). RUBubbles as a novel tool to study categorization learning.

Behavior Research Methods, 54(4), 1778–1793. https://doi.org/10.3758/s13428-021-

01695-2

Arnott, S. R., Cant, J. S., Dutton, G. N., & Goodale, M. A. (2008). Crinkling and crumpling: An

auditory fMRI study of material properties. NeuroImage, 43(2), 368–378.

https://doi.org/10.1016/j.neuroimage.2008.07.033

Banaie Boroujeni, K., Sigona, M. K., Treuting, R. L., Manuel, T. J., Caskey, C. F., &

Womelsdorf, T. (2022). Anterior cingulate cortex causally supports flexible learning

under motivationally challenging and cognitively demanding conditions. PLOS Biology,

20(9), e3001785. https://doi.org/10.1371/journal.pbio.3001785

Biederman, I., & Gerhardstein, P. C. (1993). Recognizing depth-rotated objects: Evidence and

conditions for three-dimensional viewpoint invariance. Journal of Experimental

Psychology: Human Perception and Performance, 19(6), 1162–1182.

https://doi.org/10.1037/0096-1523.19.6.1162

Boroujeni, K. B., Watson, M., & Womelsdorf, T. (2022). Gains and Losses Affect Learning

Differentially at Low and High Attentional Load. Journal of Cognitive Neuroscience,

34(10), 1952–1971. https://doi.org/10.1162/jocn_a_01885

bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Bowman, C. R., & Zeithamova, D. (2020). Training set coherence and set size effects on concept

generalization and recognition. Journal of Experimental Psychology: Learning, Memory,

and Cognition, 46(8), 1442–1464. https://doi.org/10.1037/xlm0000824

Chuang, L. L., Vuong, Q. C., & Bülthoff, H. H. (2012). Learned Non-Rigid Object Motion is a

View-Invariant Cue to Recognizing Novel Objects. Frontiers in Computational

Neuroscience, 6. https://doi.org/10.3389/fncom.2012.00026

Corriveau Lecavalier, N., Ouellet, É., Boller, B., & Belleville, S. (2020). Use of immersive

virtual reality to assess episodic memory: A validation study in older adults.

Neuropsychological Rehabilitation, 30(3), 462–480.

https://doi.org/10.1080/09602011.2018.1477684

Folstein, J. R., Gauthier, I., & Palmeri, T. J. (2012). How category learning affects object

representations: Not all morphspaces stretch alike. Journal of Experimental Psychology:

Learning, Memory, and Cognition, 38(4), 807–820. https://doi.org/10.1037/a0025836

Ghazizadeh, A., Griggs, W., & Hikosaka, O. (2016). Object-finding skill created by repeated

reward experience. Journal of Vision, 16(10), 17. https://doi.org/10.1167/16.10.17

Hassani, S. A., Neumann, A., Russell, J., Jones, C. K., & Womelsdorf, T. (2023). M1 -selective

muscarinic allosteric modulation enhances cognitive flexibility and effective salience in

nonhuman primates. Proceedings of the National Academy of Sciences, 120(18),

e2216792120. https://doi.org/10.1073/pnas.2216792120

Howett, D., Castegnaro, A., Krzywicka, K., Hagman, J., Marchment, D., Henson, R., Rio, M.,

King, J. A., Burgess, N., & Chan, D. (2019). Differentiation of mild cognitive impairment

using an entorhinal cortex-based test of virtual reality navigation. Brain, 142(6), 1751–

1766. https://doi.org/10.1093/brain/awz116
bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Kemp, L. T., Smeets, T., Jansen, A., & Houben, K. (2024). Aversive conditioning is impaired in

impulsive individuals: A study on learning asymmetries. Journal of Behavior Therapy

and Experimental Psychiatry, 83, 101939. https://doi.org/10.1016/j.jbtep.2023.101939

Knutson, A. R., Hopkins, R. O., & Squire, L. R. (2012). Visual discrimination performance,

memory, and medial temporal lobe function. Proceedings of the National Academy of

Sciences, 109(32), 13106–13111. https://doi.org/10.1073/pnas.1208876109

LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444.

https://doi.org/10.1038/nature14539

McIntire, J. P., Havig, P. R., & Geiselman, E. E. (2014). Stereoscopic 3D displays and human

performance: A comprehensive review. Displays, 35(1), 18–26.

https://doi.org/10.1016/j.displa.2013.10.004

Mercer, T., & Duffy, P. (2015). Rapid Communication: The loss of residual visual memories over

the passage of time. Quarterly Journal of Experimental Psychology, 68(2), 242–248.

https://doi.org/10.1080/17470218.2014.975256

Minda, J. P., & Smith, J. D. (2001). Prototypes in category learning: The effects of category size,

category structure, and stimulus complexity. Journal of Experimental Psychology:

Learning, Memory, and Cognition, 27(3), 775–799. https://doi.org/10.1037/0278-

7393.27.3.775

Pusch, R., Packheiser, J., Azizi, A. H., Sevincik, C. S., Rose, J., Cheng, S., Stüttgen, M. C., &

Güntürkün, O. (2023). Working memory performance is tied to stimulus complexity.

Communications Biology, 6(1), 1119. https://doi.org/10.1038/s42003-023-05486-7

Simonyan, K., & Zisserman, A. (2014). Very Deep Convolutional Networks for Large-Scale

Image Recognition (Version 6). arXiv. https://doi.org/10.48550/ARXIV.1409.1556

bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Smith, S. A. (2019). Virtual reality in episodic memory research: A review. Psychonomic Bulletin

& Review, 26(4), 1213–1237. https://doi.org/10.3758/s13423-019-01605-w

Tan, M., & Le, Q. V. (2019). EfficientNet: Rethinking Model Scaling for Convolutional Neural

Networks. https://doi.org/10.48550/ARXIV.1905.11946

Todd, J. T. (2004). The visual perception of 3D shape. Trends in Cognitive Sciences, 8(3), 115–

121. https://doi.org/10.1016/j.tics.2004.01.006

Treuting, R. L., Boroujeni, K. B., Gerrity, C. G., Tiesinga, P., & Womelsdorf, T. (2024). Anterior

Cingulate Cortex Causally Supports Meta-Learning.

https://doi.org/10.1101/2024.06.12.598723

Wallraven, C., Bülthoff, H. H., Waterkamp, S., Van Dam, L., & Gaißert, N. (2014). The eyes

grasp, the hands see: Metric category knowledge transfers between vision and touch.

Psychonomic Bulletin & Review, 21(4), 976–985. https://doi.org/10.3758/s13423-013-

0563-4

Wang, Z., Bovik, A. C., Sheikh, H. R., & Simoncelli, E. P. (2004). Image Quality Assessment:

From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing,

13(4), 600–612. https://doi.org/10.1109/TIP.2003.819861

Watson, M. R., Traczewski, N., Dunghana, S., Boroujeni, K. B., Neumann, A., Wen, X., &

Womelsdorf, T. (2023). A Multi-task Platform for Profiling Cognitive and Motivational

Constructs in Humans and Nonhuman Primates [Preprint]. Animal Behavior and

Cognition. https://doi.org/10.1101/2023.11.09.566422

Watson, M. R., Voloh, B., Naghizadeh, M., & Womelsdorf, T. (2019). Quaddles: A

multidimensional 3-D object set with parametrically controlled and customizable

bioRxiv preprint doi: https://doi.org/10.1101/2024.12.19.629479; this version posted December 20, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

features. Behavior Research Methods, 51(6), 2522–2532. https://doi.org/10.3758/s13428-

018-1097-5

Wen, X., Neumann, A., Dhungana, S., & Womelsdorf, T. (2024). Flexible Learning and Re-

ordering of Context-dependent Object Sequences in Nonhuman Primates.

https://doi.org/10.1101/2024.11.24.625056

Womelsdorf, T., Thomas, C., Neumann, A., Watson, M. R., Banaie Boroujeni, K., Hassani, S. A.,

Parker, J., & Hoffman, K. L. (2021). A Kiosk Station for the Assessment of Multiple

Cognitive Domains and Cognitive Enrichment of Monkeys. Frontiers in Behavioral

Neuroscience, 15, 721069. https://doi.org/10.3389/fnbeh.2021.721069

Womelsdorf, T., Watson, M. R., & Tiesinga, P. (2021). Learning at Variable Attentional Load

Requires Cooperation of Working Memory, Meta-learning, and Attention-augmented

Reinforcement Learning. Journal of Cognitive Neuroscience, 1–29.

https://doi.org/10.1162/jocn_a_01780

Wong, A. C.-N., Palmeri, T. J., & Gauthier, I. (2009). Conditions for Facelike Expertise With

Objects: Becoming a Ziggerin Expert—but Which Type? Psychological Science, 20(9),

1108–1117. https://doi.org/10.1111/j.1467-9280.2009.02430.x

Yazdanpanah, A., Wang, M. C., Trepka, E., Benz, M., & Soltani, A. (2024). Contributions of

statistical learning to learning from reward feedback.

https://doi.org/10.1101/2024.04.27.591445