ED5340 - Data Science: Theory apa th y
and Practise an
h ug
u t
M
a n
L29 - Deep Networks in Practise
n a t h
a
Ram
Ramanathan Muthuganapathy ([Link]
Course web page: [Link]
Moodle page: Available at [Link]
Fields / Topics that employ DL
• Computer Vision (Major one but not the only one)
h y
• Image data as the input a t
ap
n a
• Object recognition / Classification / Search & Retrieval u g
th
u
• Segmentation / Decomposition M
a n
• 3D Models - Graphics / CAD t h
n a
a
• Point cloud or Mesh model a m
R
• Classification / Search & Retrieval
• Segmentation / Decomposition
• Reconstruction
• Feature recognition (typically, mechanical model features)
2
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
Fields / Topics that employ DL
• Sketch-based modelling
h y
• Input are sketches (images!) a t
ap
an
• Classification / Search & Retrieval
h ug
u t
• Sketch clean-up M
a n
• Analysis th
n a
a
• Inputs - Points / Images Ram
• Prediction (Such as stresses / strains)
• Interpretation of data
• Flow Analysis (CFD)
• Speech / NLP / Signal processing / Medical Imaging
3
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
Point cloud
• A 3D model represented only as a set of pointshy
a t
ap
• No predetermined order (Unordered cloudgaof n points)
h u
u t
M
a n
t h
n a
a
a m
R
4
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
Mesh model
• Also called as tessellated a th y
models ap
an
h ug
t
• Triangulated data (Input file M
u
will consist of only vertices of a n
th
triangles) n a
a
R am
• No topology / geometry
information
5
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
Mesh model
h y
a t
ap
an
h ug
u t
M
a n
th
n a
a
Ram
6
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
outer boundary using a Delaunay triangulation based method [10], hole detection problem
computes a best approximation of inner boundaries of P . The proposed algorithm for hole
Reconstruction
detection is a unified one which reconstructs inner boundaries (if any) for both BS and
DP as shown in Figure 1.4.
• Input is the point cloud h y
a t
ap
• Output
1.1.2 isSurface
the mesh model
Reconstruction an
h ug
u t
M
a n
th
n a
a
R am
7
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
(a) Input point set (b) Output model
Point / shape completion
• Input - Incomplete point cloud (missing data) h y
a t
ap
• Output - Completed point cloud an
h ug
u t
M
a n
th
n a
a
Ram
8
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
Feature identification
• Input - Mesh model h y
a t
ap
• Output - Identify features like holes / slotsgaetc.
n
h u
u t
M
a n
th
n a
a
Ram
9
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
(a) (b)
Feature identification
Figure 1.4: Models having interacting features (multiple openings).
1.1.3 Part clustering
• Input -Part
Mesh model y
clustering is a process of grouping geometrically similar
h models into the same clus-
a t
ap
• Outputter- and
Group / Cluster
assigning similar
them a cluster [Link]
For example,
ga nFigure 1.5(a) shows a segmented
h u
CAD model which are then used to cluster them t
u (Figure 1.5(b)).
M
a n
t h
n a
a
a m
R
(a) (b)
Figure 1.5: (a) Segmented CAD model. (b)10Clustered segments of the same model.
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
Point cloud retrieval (roof)
h y
a t
ap
an
h ug
u t
M
a n
th
n a
a
Ram
11
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
Partial search
• Input - part of a model h y
a t
ap
• Output - Similar models consisting the input ga n as a part(s)
h u
u t
M
a n
t h
n a
a
a m
R
Figure 1.6: Example of a Part-based query and the target CAD model
12
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
Sketch processing
hand drawn sketches
• Input - Sketch of a 3D model h y
a t
ap
• Output - Similar 3D models. an
h ug
u t
• Sketch completion / sketch a n
M
clean up is also another th
n a
interesting problem a
Ram
13
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
Segmentation
h y
a t
ap
an
h ug
u t
M
a n
th
n a
a
Ram
14
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
Point set segmentation
• Input - Point cloud h y
a t
ap
• Output - classify the points an
belonging to a particular h ug
u t
segment. M
a n
th
n a
a
Ram
15
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
Reconstructing a noisy model
• Input - model with noise h y
a t
ap
• Output - Reconstructed model after removing ga n noise
h u
u t
M
a n
t h
n a
a
a m
R
16
Ramanathan Muthuganapathy, Department of Engineering Design, IIT Madras
DL-based Analysis of Loading conditions
• FEA (Finite
element
analysis) is a
popular.
• DL-based
approach
towards
analysis.