Cinematographic and Geometric Criteria for Virtual Camera Path

Generation for the Visualization of Shipwreck Data

Katerine Davis

, Vaibhav K. Viswanathan

Christopher M. Clark

, Timothy Gambin

and Zo

e J. Wood

Cal Poly, San Luis Obispo, U.S.A.

Harvey Mudd College, Claremont, U.S.A.

University of Malta, Msida MSD 2080, Malta

Keywords:

Virtual Camera Path, Probabilistic Road Map, Rule of Thirds, Marine Archeology Visualization.

Abstract:

This paper presents the use of cinematographic and geometric measures for determining the path of a virtual

camera for the generation of computer graphics video sequences focused on showing an underwater shipwreck.

This work rises from the applied challenge of mapping underwater shipwrecks, reconstructing a computer

graphics model then creating an educational visualization of the wreck. The primary algorithm presented

in this work computes the optimal camera pitch and height along a path utilizing a probabilistic roadmap

(PRM) that weights nodes using a computational models of cinematographic and geometric principles. These

principles were used to evaluate potential viewpoints and inﬂuence whether or not a view is used in the ﬁnal

path. Speciﬁcally, the algorithm uses a computational evaluation of the cinematographic ‘rule of thirds’ and

a geometric evaluation of the model normals relative to the camera viewpoint. A user study of video output

from the system suggests that our computed paths are ranked higher than simple circular or random paths and

that the ‘rule of thirds’ is a more important criteria than the geometric principle explored.

1 INTRODUCTION

Determining the path of a virtual camera through a

digital scene in order to display the most relevant in-

formation in an aesthetically appealing way is an es-

sential challenge when producing video output for a

computer graphics applications. There exist millions

of exquisite computer graphics models and scenes

carefully crafted by artists and computer scientists.

In order to show off these models and scenes, typ-

ically animators must hand create camera paths, of-

ten called ﬂy-throughs. They do this by carefully

choosing camera viewpoints that display speciﬁc fea-

tures of the models and interpolating smooth paths be-

tween these points. This is a time consuming process

that requires artistic ability to do well. We present a

method to generate automatic ﬂy-throughs for three-

dimensional (3D) computer graphics scenes focused

on a single model, building off the rich history of cin-

ematographic principles, speciﬁcally focused on the

general composition ‘rule of thirds’ (Krages, 2005).

We use this cinematographic and an additional geo-

metric consideration to generate videos that are ap-

pealing to users and feature good views of the most

salient information in our scene.

In particular this work is focused on creating edu-

cational videos that allow a user to see and understand

shipwrecks of archaeological signiﬁcance. This work

is a part of a larger project (von Fock et al., 2017) that

uses underwater robots, sonar scans, video footage,

and photogrammetry to create 3D computer graphics

models of shipwrecks. Speciﬁcally, this work focuses

on generating camera paths in a virtual underwater

world, featuring the reconstruction of the World War

II shipwreck, X127, mapped by the 2016 ICEX team.

Arising from this multi-disciplinary robotics and

computer graphics project, our virtual camera work

is inspired by the related problem of determining the

path of the robot used in the same project. Speciﬁ-

cally, to create a computer graphics reconstruction of

the real world ship wrecks we use a photogramme-

try pipeline that uses discrete frames from video data

acquired by Autonomous Underwater Vehicle (AUV)

deployments. To determine the path of the robot

for optimal video coverage for the photogrammetry

pipeline, we use a probabilistic roadmap (PRM) algo-

rithm that utilizes a rapidly-exploring random tree to

quickly cover a space and generate small maps with

Davis, K., Viswanathan, V., Clark, C., Gambin, T. and Wood, Z.

Cinematographic and Geometric Criteria for Virtual Camera Path Generation for the Visualization of Shipwreck Data.

DOI: 10.5220/0006543801910198

In Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2018) - Volume 1: GRAPP, pages

191-198

ISBN: 978-989-758-287-5

191

good coverage. PRMs are ideal for classic robotics

problems because they are highly customizable and

can create paths that follow kinematic constraints of

different robots. PRMs have also been used in a vari-

ety of other applications such as video games, charac-

ter animation, and even computational biology (Hsu

et al., 2003). For determining the path of the robot,

a measure of information gain tied with the gradient

of an occupancy map generated via sonar mapping

is used to determine an optimal path for acquiring

video footage of the wreck (Viswanathan et al., 2017).

The video footage is then used in a photogrammetry

pipeline to create a computer graphics model of the

wreck.

Similarly, our algorithm to compute the path of the

virtual camera uses a PRM planning algorithm to gen-

erate the ﬂy-through camera path, given that PRMs

have good coverage while keeping maps small. In or-

der to evaluate a given node in our camera path, we

use both cinematographic and geometric considera-

tions to determine the value of the node. Speciﬁcally,

these considerations evaluate viewpoints from heights

and camera pitches generated along a speciﬁed path.

The “rule of thirds” is used to characterize cinemato-

graphic principles, and object normals are representa-

tive of geometric evaluation.

This paper presents: our image processing system

that renders a given 3D scene from a speciﬁed camera

position and direction, then analyzes the image using

computer vision with regards to cinematographic and

geometric principles; our probabilistic roadmap algo-

rithm that generates customizable, visually appealing

camera paths for a ﬂy-through of a 3D scene focused

on a single model. We also present our evaluation of

variations in the path selection criteria via user’s eval-

uations of videos produced by our system. In general,

users are pleased with the paths our system creates

and prefer the paths generated with cinematographic

principles over simple, ﬂat paths.

2 RELATED WORKS

A large body of work exists related to probabilistic

roadmaps, virtual cameras, and computational evalu-

ation of image composition.

Cinematography and Image Composition. A

canonical work in the area of combining cinematog-

raphy with computer science is that of Christianson

et al. (Christianson et al., 1996). This paper created a

camera control language to formalize cinematography

principles for use in computer graphics. The idioms

encoded focus on character interaction and movement

rather than composition rules such as the rule of thirds

used in our work.

The detection of the rule of thirds in still images is

presented in (Mai et al., 2011). Mai et al. use saliency

and generic objectness analysis to determine the sub-

ject of a photo and analyze where it is relative to the

thirds lines. Other works detecting the rule of thirds

utilize it along with other image composition rules to

create aesthetically pleasing images through cropping

and retargeting (Liu et al., 2010), (Setlur et al., 2005).

Since our work focuses on scenes with a single main

object we use a simple object detection technique to

evaluate the rule of thirds.

Recently, evidence of the desirability of an auto-

matic virtual camera planning, Unity Engine incorpo-

rated a virtual camera system, Cinemachine, to en-

able users with several virtual camera planning op-

tions (cin, 2017). The system includes the ability to

specify a camera to track AI characters in a game,

identify ‘priority based shots’ and other features.

Motion Planning Algorithms for Camera Control.

Recent work by Joubert, et. al., (Joubert et al., 2016),

focuses on taking well-composed video footage from

a quadcopter of human subjects interacting with each

other using GPS and Inertia Measurement Unit sen-

sors. To transition between static shots, the drone uses

a real-time trajectory planning algorithm that tries to

maintain visually pleasing shots while moving.

Another related work is that of Li and Cheng who

navigate virtual environments with a real-time camera

control module (Li and Cheng, 2008). This module

generates camera motions that automatically control

a third-person camera for use in video games. They

maintain a real-time camera by using a lazy PRM that

follows the game avatar. The PRM uses intercuts to

avoid occlusion and collisions while preserving cine-

matography rules.

Other research related to automatic camera con-

trols include visualizations focused on historical data

(Stoev and Straer, 2002), examinations of viewpoint

quality (Sokolov and Plemenos, 2005), and surveys

of camera controls primarily applied to game settings

(Christie and Olivier, 2009).

Probabilistic Roadmaps. A vast body of work ex-

ists examining probabilistic roadmap theory and its

applications in robotics (Li and Shie, 2002), includ-

ing those using RRTs and PRMs for planning the

path of an AUV, (Poppinga et al., 2011), (Lavalle,

1998), (Lavalle et al., 2000), (Candeloro et al., 2015).

The robotics portion of our shipwreck mapping

project uses a PRM planner and an AUV to help cre-

ate the shipwreck model (Viswanathan et al., 2017).

GRAPP 2018 - International Conference on Computer Graphics Theory and Applications

192

An occupancy map is created from a high level AUV

scan of the shipwreck area and used in a PRM. The

PRM creates additional AUV missions that make low

altitude ﬂyovers of the area optimized for “maximiz-

ing information gain” or getting good video data of

the shipwreck. The photogrammetry pipeline (von

Fock et al., 2017) used to create the model of the ship-

wreck requires photos of the shipwreck from as many

angles as possible. Our “high weight node” list was

inspired by this work’s PRM that maximizes infor-

mation gain. However, this work is optimizing for a

high number of camera angles to use photogramme-

try while we focus on maximizing viewpoints from a

cinematographic and geometric perspective.

3 ALGORITHM

We present a system to generate ﬂy-through anima-

tions using a robotics motion planning algorithm,

probabilistic roadmaps. Cinematographic and geo-

metric principles, speciﬁcally the rule of thirds and

model normals, are used to generate virtual camera

paths, resulting in visually pleasing video output. The

sample scene used here features a digital 3D model of

an actual shipwreck in Malta, the X127, placed in an

underwater scene. The basic algorithm is as follows:

R(N, E) = Roadmap(Nodes, Edges);

Data: rootNode, pathLength

Result: Path of pathLength from rootNode

highWeightNodes[];

while currentPathLength < pathLength do

// select node c to expand from

if even iteration of loop then

c ← highWeightNodes[iteration];

end

else

c ← rand node in R;

while c < weightThreshold;

end

randomly generate c’ from c;

calculate edge e from c to c’;

R.add(c’, e);

currentPathLength ← c’.pathLength;

c’.weight ← calculateWeight(c’);

if c’.weight > highWeightThreshold then

highWeightNodes.add(c’);

end

Algorithm 1: High Level PRM Camera Path Genera-

tion Algorithm.

Figure 1: Top down view of the the shipwreck scene. Semi-

elliptical path of varying radius shown in pink. Caustics

turned off for clarity.

3.1 PRM Algorithm

PRMs use a RRT to explore a conﬁguration space

with small maps that provide good coverage. The

conﬁguration space used in this work is the height and

pitch of the camera. At this time, the camera paths

are generated along a set elliptical path with variation

in radius. An example path shown from a top down

view can be seen in Figure 1. The underwater world

has various rock and seaweed models throughout and

features a central shipwreck model that is our primary

target for the camera to view from interesting angles.

Nodes are placed on a tree starting from a single

root node, which is the camera’s starting conﬁgura-

tion. An entire tree of nodes is called a roadmap. This

roadmap contains all nodes in one run of the PRM al-

gorithm, regardless of whether or not these nodes end

up on the ﬁnal ﬂy-through path. We use the term path

to describe a set of nodes on our roadmap that form

a complete course from start to end conﬁguration of

the desired length. Broadly, the algorithm adds nodes

to a roadmap until a path of the desired node count

is generated. First, a starting conﬁguration is selected

with a good viewpoint. Then, a node to expand from

is selected using speciﬁc criteria and new nodes are

generated again with speciﬁc criteria discussed below.

A node’s value is calculated based on the measured

quality of the rendered view from that node.

Root Node Generation. All paths and roadmaps

must begin with a starting conﬁguration, speciﬁed in a

root node. To start from a good initial conﬁguration,

200 potential root nodes are generated by selecting

random heights and pitches within reasonable start

values. The weight of all these nodes are calculated

based on the current viewpoint evaluation criteria as

Cinematographic and Geometric Criteria for Virtual Camera Path Generation for the Visualization of Shipwreck Data

193

described below in Section 3.2 and the best node is se-

lected as the root node. Once the root node is chosen,

it is added to the roadmap and the algorithm proceeds

with the node selection step.

Node Selection. For the algorithm to proceed, a

node to expand from must be selected. We want to

select a node with a better weight to generate more

aesthetically pleasing paths. Throughout the entire

path generation process, a list of “high weight” nodes

is maintained that have values above a certain thresh-

old. On half of the iterations through the path genera-

tion, a node from this list is selected to expand from.

On the alternate loop iterations a random node with a

weight above a threshold value in the roadmap is se-

lected to expand from. This threshold is signiﬁcantly

lower than the threshold for the high weight node list.

Additionally, this threshold decreases slightly each it-

eration in order that a node will be selected even if

the roadmap only contains low quality nodes. Using

node weight thresholds can be seen as a pruning step

and this process helps keep the roadmap size small

by not expanding the roadmap in directions with low

quality nodes.

Node Generation. In order to display our scene

in a compelling way, the roadmap must expand out

through our conﬁguration space to show the ship-

wreck at a variety of angles. Once a node to ex-

pand from has been selected as described above, a

new node is generated based off this old node. This

node generation is done by taking the height of the old

node and adding and subtracting a height delta. A ran-

dom height is then chosen within this new range. The

same process happens for pitch, selecting a random

pitch within a delta of the old node’s pitch. Height

delta was typically in the range of 1-3 and pitch delta

was generally between 5

◦

and 15

◦

. In our current vir-

tual world, one meter in the real world is about 1.6

units in our virtual world.

Given that in general, ﬂy-throughs follow a mostly

circular path, we use a simpliﬁcation on the planning

of the camera and determine the x and z coordinates

along with the yaw and roll by calculating the polar

coordinates of a circle around the main model. We

allow for variations in the radius (25-30 units) and ec-

centricity to create interesting paths. Figure 1, shows

the top down view of one variation on a path gener-

ated by our system. In addition, during node genera-

tion, the height is constrained to be above the seaﬂoor

and below a speciﬁed maximum height. Once a new

node is generated, we must evaluate the scene from

that node’s position and camera view direction.

3.2 Viewpoint Evaluation

The process of evaluating the camera viewpoint spec-

iﬁed in a given node is based on cinematographic and

geometric criteria. We choose to evaluate these crite-

ria in the image domain, thus must render our com-

puter graphics scene to evaluate the image. The node

object is passed from the PRM generation code to

the OpenGL code, rendered using OpenGL with the

results then evaluated using OpenCV. The calculated

weight is then set for the node and used in the node

selection process for the next iteration of the PRM

loop.

The 3D underwater OpenGL scene includes sand,

rocks, seaweed, bubbles, caustics, and the X127 ship-

wreck model. An example of all these elements in

the scene can be seen in Figure 5. The primary ele-

ment in the scene is the X127 shipwreck, created from

an actual shipwreck in Malta using a photogramme-

try pipeline using input video gathered from an AUV.

Since the X127 wreck lays on a diagonal, the deeper

part of the wreck sits in water that is too murky for

the AUV to take high quality video footage.

A PRM node’s position and direction vectors are

used to set a virtual camera in the OpenGL scene.

Then the resulting scene is written to a framebuffer

and transferred to the OpenCV image format, which

is used to evaluate how closely the image follows the

rule of thirds and whether the model is primarily fac-

ing towards the camera.

Rule of Thirds Detection. One of the canonical

rules of aesthetics from a cinematographic perspec-

tive is the rule of thirds, which can be described as

splitting an image up into three uniform sized sec-

tions, both horizontally and vertically. The subject

of the photo or horizon line should be placed along

one of these lines or at the intersection of two such

lines as can be seen in many famous photographs and

paintings. Thus for automatic camera path genera-

tion a good view of the scene should follow the rule

of thirds so our algorithm evaluates nodes relative to

how well a given view follows this rule.

We combine several of the methods provided in

OpenCV to evaluate how well an image conforms to

the rule of thirds. First the input image is converted to

grayscale and blurred. The image is then thresholded

to detect edges and then the contours or outlines of

the shapes in the image are computed. The contours

are approximated as polygons and bounding rectan-

gles and circles are formed. An example of the output

from these computations can be seen in Figure 2.

To use this methodology in our program, a ker-

nel size of ﬁve by ﬁve is used in the normalized box

GRAPP 2018 - International Conference on Computer Graphics Theory and Applications

194

Figure 2: OpenCV methodology to draw bounding rectan-

gles and circles around detected contours in an image.

Figure 3: Rule of thirds detection. On the right the OpenGL

scene is rendered and on the left, OpenCV is used to detect

the largest shape in the image and compared to the thirds

lines of the image.

ﬁlter. In addition, bounding circles are eliminated

and only one bounding rectangle is preserved, specif-

ically the largest. There are various ways to consider

a rule of thirds composition, thus we conducted a user

study with static images of our scene and found that

users preferred the top horizontal line of the ship to be

aligned with top third horizontal extent of the image,

i.e. where the shipwreck created a horizon line at the

top thirds line of the image. Thus, the largest y coor-

dinate of the largest bounding rectangle is found and

used to compare against the top third of the entire im-

age to evaluate whether the key feature of our scene

conforms to the rule of thirds as shown in Figure 3.

See Equation 1. This measure is set as the weight of

the node to be used in node selection. A limitation

of our method is that we currently can only have one

main model in our scene for our detection algorithm

to work well.

weight = |(BBox.y/ImageHeight − 2/3)|

(1)

Model Normals Detection. From a geometric per-

spective, we want camera views that show us as much

of the model as possible. The node weight calcula-

tion detects how much of a given object is pointing

towards the camera via use of the model’s normals.

Speciﬁcally, a vertex shader is used to transform the

normals into the camera’s eye space. When these nor-

mals are then used to color the rasterized fragments, a

normal that is pointing primarily in the direction of

the camera will be colored blue. OpenCV is used

to detect the proportion of pixels that represent frag-

ments facing the camera, which is then used to weight

the nodes in the search space.

Thirds and Normals Combination. The rule of

thirds and normal detection both have shortcomings

so we explored combining them to try to get the best

path possible. The rule of thirds is optimal for creat-

ing smooth paths but can occasionally score well for

views of the wreck from poor angles. For example, a

very low viewpoint, almost from the seaﬂoor can cre-

ate a path following the rule of thirds just as well as

one from a higher, more scenic view. The normal de-

tection does a much better job with this, but can create

very jumpy paths that leap from one extreme angle to

another. To combine the thirds scores that are better

if they are lower and normals scores where a higher

score is better, we use the following equation:

weight = (1 −|(BBox.y/ImageHeight − 2/3)|)+

(

ImageSize

∑

n=0

pixel.blue > threshold)/ImageSize

(2)

An example of a node viewpoint being evaluated with

both methods is shown in Figure 4. This combina-

tion method was noticeably slower because it required

rendering the scene, converting the image to OpenCV

format, and running OpenCV algorithms twice for

each node viewpoint.

Node Weight & Best Weight List. To drive the

path towards a reasonable solution in a shorter amount

of time, a highWeightNode list was used during the

node selection process and every other iteration of

the PRM loop, a node is selected from this list. For

the rule of thirds, nodes with weight 0.04 or lower

were added to the best weight list while nodes with

weight 0.25 or higher were considered good for nor-

mals scores. Since these methods were combined by

subtracting the thirds score from 1 to create positive

values, a good combination score was deﬁned as 1.21.

Storing Complete Paths. Each time a new node is

added to the roadmap, the algorithm checks if that

Cinematographic and Geometric Criteria for Virtual Camera Path Generation for the Visualization of Shipwreck Data

195

Figure 4: Combination node weight calculation. Rule of

thirds and model normals are detected for each viewpoint

and the combined weights are set as the node weight.

node completes a path of the desired path length. If a

full length path is found then the roadmap is traversed

from the newest node back up to the root node. Posi-

tion and direction vectors of each node in the path are

then written out to a ﬁle allowing our system to play-

back a path from a ﬁle. Videos can then be created by

using the full list of camera positions and directions

to create a Catmull-Rom spline between the nodes. A

spline path interpolating between nodes in a generated

path can be seen in pink in Figure 1.

4 RESULTS

In this work, a PRM algorithm adds nodes to a tree

until a path of sufﬁcient node count has been gener-

ated. A node to expand from is selected, then new

nodes are generated given precise parameters. Vir-

tual camera viewpoints, speciﬁed in new nodes, are

evaluated on cinematographic and geometric princi-

ples, i.e., the rule of thirds and model normals, and

assigned a weight. These weights are used in the

node selection process, generating aesthetically pleas-

ing camera paths.

Parameters to control the path generation include

variations on the path such as length, radius, elliptical

factor, radius variation, camera speed (via spline com-

putation between nodes) and camera speciﬁc vari-

ables such as initial height and pitch, height and pitch

delta, and ﬁnally evaluation criteria such as color and

blur threshold factors. PRM variables include “high

weight” list threshold for thirds, normals, and com-

bination, frequency of selecting from high weight list

versus random selection, threshold when selecting a

random node to expand from for thirds, normals, and

combination, and limits when changing the threshold

for selecting a random node.

Table 1: Viewpoint evaluation method study results. 1 is

most preferred and 5 is least preferred. Columns 2 - 6 show

the number of participants that gave that ranking and the

last column shows the average score for each video. Video

A was the highest ranked with an average score of 2.14.

Video 1 2 3 4 5 Avg.

A (Thirds) 13 10 8 2 2 2.14

B (Random) 0 0 2 11 22 4.57

C (Combo) 10 5 9 8 3 2.69

D (Flat) 7 11 10 5 2 2.54

E (Normals) 5 9 6 9 6 3.06

4.1 User Studies

With so many different parameters we evaluated our

results with a user study comparing output videos.

Output videos include ones that keep all parameters

the same and vary the method used for calculating

node weights (the rule of thirds, model normals, com-

bination), judged against a completely ﬂat path, and

a completely random path. A second study included

videos that used a combination of thirds and normals

to calculate node weights but then varies path param-

eters, including path length, radius, deltas, elliptical

factor of path, and more. The camera viewpoints in

the best generated path from both studies can be seen

in Figure 5.

In each survey, participants were asked to rank ﬁve

15 second videos in order of preference.

Viewpoint Evaluation Methods Study. The ﬁrst

study focused on evaluating camera paths that were

generated via different evaluation methods. For this

survey, path parameters were kept constant and only

node evaluation criteria were changed. The paths had

a slightly elliptical orbit that was 0.75 width in the z

component. The initial height varied between 4 and

12 and the initial pitch angle varied between 45

◦

and

◦

below the horizon line. The max height delta pos-

sible was 2.5 and the max pitch change was 7.5

◦

. The

thresholds of what deﬁned a “good” node were kept

constant throughout all path generation for both sur-

veys. The viewpoint evaluation methods used to pro-

duce each video are as follows: Video D was a com-

pletely ﬂat path with no variation in height, Video B

displayed a path that was randomly generated with the

above parameters, Videos A, E, and C used the rule

of thirds, model normals, and a combination of both

respectively to evaluate potential camera viewpoints.

To avoid bias, 30 paths were generated of each

type and the path with the best average node score

was chosen for the study. Results of the study for the

35 participants are displayed in Table 1.

GRAPP 2018 - International Conference on Computer Graphics Theory and Applications

196

The results show that people preferred the path

generated with only the rule of thirds. The second

highest path was ﬂat, followed closely by the combi-

nation path. Normals came in a solid fourth with ran-

dom at a strong last place. The results were surpris-

ingly spread out, with all paths except random having

at least two rankings for every possible position.

Path Types Study. This study kept the viewpoint

evaluation method constant, using a combination of

rule of thirds and model normals. The variable being

studied here was path type. Videos A through E re-

spectively displayed paths with large variation, many

nodes, small variation and perfectly circular, largely

elliptical, and random. Excluding random, 30 paths

of each type were generated and the best scoring path

of each type was used in the study. Since the random

path used no evaluation criteria, one random path was

generated and used in the study. This study had 32

participants total. The primary conclusion from this

study was that user’s preferred the automatic camera

paths and in particular the paths with the largest al-

lowable variations in view and path radius variations,

with 18 of 32 people ranking it number 1 and 12 peo-

ple ranking it number 2.

Figure 5: Camera viewpoints used in the winning path in

the user studies. A spline interpolates between these views

to create a 15 second ﬂy-through of the scene.

From these studies, we can conclude that overall

the use of our automatic camera path generation is

appealing to users, winning out over planar circular

paths and random paths and with the rule of thirds

being more important than our current geometric cri-

teria of using normals alone. There is a large variety

of variations and future work discussed below.

Table 2: Information corresponding to best paths; average

weight of all nodes in ﬁnal path, number of nodes in the

entire roadmap and seconds spent generating the roadmap.

Criteria Num Avg Roadmap Time

nodes Weight Size

Thirds 12 0.0131 1134 38.5

Norms. 12 0.1681 202 6.4

both 12 1.0467 663 36.4

both 12 0.99642 243 19.5

both 45 0.9858 1080 48.8

both 10 0.993663 56 12.4

both 12 0.917036 238 19.4

4.2 Performance and Limitations

Implementation of this work was done in C++ us-

ing the Open Graphics Library application program-

ming interface and the Open Computer Vision library.

It was run on a 2015 MacBook Pro running macOS

Sierra 10.12.3 with a 2.9 GHz Intel Core i5 processor.

All motion planning algorithms can be time and

memory consuming to run, with the PRM being no

exception. One way to generate a smooth camera

path is to put so many nodes on a circular path that

each viewpoint is extremely close together. The cam-

era will ﬂow smoothly through node transitions due to

the short distance between viewpoints. We attempted

this in our initial implementation, generating paths of

length 40 in about 10 seconds with full roadmaps con-

taining over 20,000,000 nodes. The time it took to

generate longer paths went up exponentially, taking

30 seconds to generate a path of length 50 and a path

of length 60 could not be generated in ﬁve minutes.

We experimented with many different optimizations

such as limiting the number of children per node, spa-

tially hashing the conﬁguration space and limiting the

number of nodes in each spatial range, and selecting a

random point in the conﬁguration space and expand-

ing off the nearest node to it.

Using splines to interpolate the camera path be-

tween viewpoints, enabled us to use fewer nodes in

a path, which in turn meant lower time and memory

costs, allowing us to generate and compare a larger

number of paths. The ﬁnal algorithm uses around

12 nodes to generate a path. This decision is justi-

ﬁed by the assumption that a movie director creating

an overview scene like our circular ﬂy-through would

only place a few key viewpoints and let the camera

operator move smoothly between these views. The

evaluation type, path length, average node weight,

roadmap size, and time to generate the path for the

paths evaluated in our user study, excluding the ﬂat

and random ones, are summarized in Table 2.

Cinematographic and Geometric Criteria for Virtual Camera Path Generation for the Visualization of Shipwreck Data

197

Limitations and Future Work. Our ﬂy-through

generation system has a number of limitations that

will be improved in future iterations of this project.

First, our system currently only works on scenes with

one central model because our rule of thirds detection

algorithm only looks at the single largest shape. A

more complete comparison with existing work would

also be highly valuable for positioning this work in a

larger context. We are exploring adding more motion

modeling constraints on the camera motion in terms

of pitch and height and adding other variables that can

be varied and computed using the PRM. These vari-

ables notably include completely free positioning of

the camera and solving for the yaw of the camera. We

are also exploring additional evaluation criteria such

as object detection of speciﬁc items of interest to the

viewer and including camera collision detection.

ACKNOWLEDGEMENTS

We acknowledge the hard work on the entire ICEX

2016 team.

REFERENCES

(2017). Cinemachine camera tools in unity engine.

https://blogs.unity3d.com/2017/01/23/cinemachine-

base-rig-now-available-for-free-creator-adam-myhill-

joins-unity/.

Candeloro, M., Mosciaro, F., Srensen, A. J., Ippoliti, G.,

and Ludvigsen, M. (2015). Sensor-based autonomous

path-planner for sea-bottom exploration and mosaick-

ing. In IFAC Conference on Manoeuvring and Control

of Marine Craft, pages 31–36.

Christianson, D. B., Anderson, S. E., and wei He, L. (1996).

Declarative camera control for automatic cinematog-

raphy. American Association for Artiﬁcial Intelli-

gence.

Christie, M. and Olivier, P. (2009). Camera control in com-

puter graphics: Models, techniques and applications.

ACM Siggraph Asia Courses.

Hsu, D., Jiang, T., Reif, J., and Sun, Z. (2003). The

bridge test for sampling narrow passages with prob-

abilistic roadmap planners. In 2003 IEEE Interna-

tional Conference on Robotics and Automation (Cat.

No.03CH37422), volume 3, pages 4420–4426 vol.3.

Joubert, N., E, J. L., Goldman, D. B., Berthouzoz, F.,

Roberts, M., Landay, J. A., and Hanrahan, P. (2016).

Towards a Drone Cinematographer: Guiding Quadro-

tor Cameras using Visual Composition Principles.

ArXiv e-prints.

Krages, B. P. (2005). Photography: The Art of Composi-

tion. Allworth Press.

Lavalle, S. M. (1998). Rapidly-exploring random trees: A

new tool for path planning. Technical report.

Lavalle, S. M., Kuffner, J. J., and Jr. (2000). Rapidly-

exploring random trees: Progress and prospects. In

Algorithmic and Computational Robotics: New Direc-

tions, pages 293–308.

Li, T.-Y. and Cheng, C.-C. (2008). Real-Time Camera Plan-

ning for Navigation in Virtual Environments, pages

118–129. Springer Berlin Heidelberg, Berlin, Heidel-

berg.

Li, T.-Y. and Shie, Y.-C. (2002). An incremental learn-

ing approach to motion planning with roadmap

management. In Proceedings 2002 IEEE Interna-

tional Conference on Robotics and Automation (Cat.

No.02CH37292), volume 4, pages 3411–3416 vol.4.

Liu, L., Chen, R., Wolf, L., and Cohen-Or, D. (2010). Op-

timizing photo composition. Computer Graphics Fo-

rum, 29(2):469–478.

Mai, L., Le, H., Niu, Y., and Liu, F. (2011). Rule of thirds

detection from photograph. In Proceedings of the

2011 IEEE International Symposium on Multimedia,

ISM ’11, pages 91–96, Washington, DC, USA. IEEE

Computer Society.

Poppinga, J., Birk, A., Pathak, K., and Vaskevicius, N.

(2011). Fast 6-dof path planning for autonomous un-

derwater vehicles (auv) based on 3d plane mapping.

In IEEE International Symposium on Safety, Security,

and Rescue Robotics (SSRR), pages 1–6. IEEE Press,

IEEE Press.

Setlur, V., Takagi, S., Raskar, R., Gleicher, M., and Gooch,

B. (2005). Automatic image retargeting. In Proceed-

ings of the 4th International Conference on Mobile

and Ubiquitous Multimedia, MUM ’05, pages 59–68,

New York, NY, USA. ACM.

Sokolov, D. and Plemenos, D. (2005). Viewpoint quality

and scene understanding. Proceedings of the Interna-

tional Conference on Virtual Reality, Archaeology and

Intelligent Cultural Heritage.

Stoev, S. L. and Straer, W. (2002). A case study on au-

tomatic camera placement and motion for visualizing

historical data. Proceedings of the Conference on Vi-

sualization.

Viswanathan, V. K., Lobo, Z., Lupanow, J., von Fock,

S. M. T. S., Wood, Z., Gambin, T., and Clark, C.

(2017). Auv motion-planning for photogrammetric

reconstruction of marine archaeological sites. In Pro-

ceedings of the 2017 IEEE International Conference

on Robotics and Automation.

von Fock, S. M. T. S., Bilich, S., Davis, K., Viswanthan,

V. K., Lobo, Z., Lupanow, J., Clark, C., Gambin, T.,

and Wood, Z. (2017). Pipeline for reconstruction and

visualization of underwater archaeology sites using

photogrammetry. In Proceedings of the 2017 ISCA

International Conference on Computers and Their Ap-

plications.

GRAPP 2018 - International Conference on Computer Graphics Theory and Applications

198