1) The document presents a method for improving BIM search engines by exploiting relationships between objects in 3D models. It uses graph theory to represent models as nodes and connections.
2) The method was tested on a Revit model containing 7,000 3D objects and 20,000 information objects. Different relevance measures were compared that took into account single object relevance, related objects, and neighbor objects.
3) For single keyword queries, the basic relevance measure achieved a maximum recall of 0.08 with perfect precision. For multiple keyword queries, maximum recall was 0.567 and average precision was 0.871. The expanded relevance measures did not significantly change search results due to the limited output of 250 items.
[3DIR] BIM Search Engine: Exploiting Interrelations between Objects when Assessing Relevance
1. Dr Peter Demian
Reader in Building Information Management
ICCCBE 2018, Tampere, Finland, June 2018
BIM Search Engine: Exploiting
Interrelations between Objects
when Assessing Relevance
2. Outline
• Introduction to 3DIR project
• Related research
• Graph Theory for studying 3D models
• Method: Relevance formulations
• Results
• Conclusions
3. Problem addressed by 3DIR:
Finding information
• Formulate query
• Identify relevant
information from
index
• Present a ranked list
of search results
4. …but if our information is linked to a 3D
artefact (…BIM)
There might be a better
way to:
• Formulate queries
• Identify relevant
information
• Present search results
5. Related Research
• BIM/CAD: more information in models
• Information Retrieval
• Topology, Graph Theory
• Literature reviewed in paper
Important because we wish to exploit
interrelations
6. Graph Theory for studying 3D models
V3D
1 Roof
2 Door
3 Wall
Vi
4 Name: Red roof
Type: Roof
5 Name: West door
Type: Door
Material: Glass
6 Name: South wall
Type: External Wall
Material: Concrete
7 This document is a
reinforcement
schedule for the South
wall (external).
En
a 1,4
b 2,5
c 3,6
d 3,7
Et
e 1,2 Touching
f 1,3 Intersecting
g 2,3 Hosting
Roof
Wall
Door
Reinforcement
schedule for wall
4
2
3
1
5
6 7
a b
c d
e
f g
7. Method 1/2
• Revit model from industry partner
• Ground+three-floor office
• 7k 3D objects, 20k “info” objects
• Test queries: single keyword or multiple keyword
• Relevant items for each query identified by human
expert
• Measures of Recall and Precision used to assess
system’s retrieval performance
• Holistic/contextual search relevance measures take
account of related items (other properties of that 3D
object, related 3D objects or “neighbours”)
8. Method 2/2: Relevance Measures
Name Equation Rationale
“Vi” Relevance S(V3D) Standard Vi Lucene
score
“Vi+V3D” Relevance C1S(V3D) + C2S(V3D) Also accounting for
relevance of 3D object
as a whole
“Vi+V3D+N”
Relevance
C3S(V3D) + C4S(V3D) +
C5S(V3D-N)
Also accounting for
relevance of
Neighbours
“Vi+V3D+N+NN”
Relevance
C6S(V3D) + C7S(V3D) +
C8S(V3D-N) + C9S(V3D-
NN)
Also accounting for
relevance of Neighbours-
of-Neighbours
9. Results 1/2: Single Keyword Queries
Query → Query 1a Query 1b Query 1c
Query Terms glazing glazed glaz*
Relevant Vi items (according to
human expert)
9 3092 3101
Vi items retrieved by 3DIR 8 250 (3DIR maximum) 250 (3DIR maximum)
“Vi” Relevance performance 3DIR successfully retrieved 8 of the
9 relevant items. The precision was
1 at all recall levels.
3DIR has a maximum of 250 search
hits, which means the maximum
possible recall is 0.08, and this was
achieved using this basic relevance
measure. Precision was 1 at all
levels.
As expected, the set of relevant
items for this query is the union of
the relevant sets for Queries 1a and
1b. The results were roughly the
same as for Query 1b.
“Vi+V3D” Relevance performance The ranking of search hits did not
change from above.
Although there were minor
differences to the items retrieved
and their rankings, the maximum of
250 search hits and the large
number of relevant items meant that
maximum precision was still 0.08,
again with no irrelevant items
retrieved.
Roughly the same as for Q1b.
“Vi+V3D+N” Relevance
performance
The ranking of search hits did not
change from above.
Same as above: slightly different
search hits and ranking, but no
change in recall and perfect
precision.
Roughly the same as for Q1b.
“Vi+V3D+N+NN” Relevance
performance
The ranking of search hits did not
change from above.
Same as above: slightly different
search hits and ranking, but no
change in recall and perfect
precision.
Roughly the same as for Q1b.
10. Results 2/2: Multiple Keyword Queries
Query 2
Query Terms internal wall door glaz*
Relevant Vi items
(according to
human expert)
238
Items retrieved 250 (3DIR maximum)
Maximum Recall 0.567
Average Precision
(averaged over 250
retrieved search
hits)
0.871
Relevance
Measure →
Performance
Criterion ↓
“Vi”
Relevance
“Vi+V3D”
Relevance
“Vi+V3D+N”
Relevance
“Vi+V3D+N+NN”
Relevance
Top Rank of
Irrelevant
Retrieved Search
Hit
134 134 134 134
Bottom Rank of
Relevant Retrieved
Search Hit
150 250 250 250
11. Conclusions
• 3DIR imposed a limit of maximum 250 search hits, which
obscured results
• Innovation presented here did not affect retrieval of
results, only inking
• Useful effect of scattering relevance measures
• Measures of Recall and Precision not sensitive enough
to measure benefit or our proposes
• Graph theoretic formulation is a useful theoretical lens
for studying and developing BIM search engines
12. THANK YOU
Peter Demian P.Demian@lboro.ac.uk
3DIR project website: http://www.3dir.org/
Free 3DIR add-in-in for Revit available from the Autodesk App Store