VISVESVARAYA TECHNOLOGICAL UNIVERSITY
Jnana Sangama, Machhe, Belagavi, Karnataka 590018
SEMINAR REPORT
on
“Dynamic Graph CNN for Learning on Point Clouds”
Submitted in partial fulfillment of the requirement
for the award of the degree of
Bachelor of Engineering
in
Information Science and Engineering
by
Bharath S[1BG17IS008]
Under the Guidance of
Dr. S Srividhya
Associate Professor
Department of Information Science and Engineering
B.N.M. Institute of Technology
Approved by AICTE, Affiliated to VTU, Accredited as grade A Institution by NAAC.
All UG branches – CSE, ECE, EEE, ISE & Mech.E accredited by NBA for academic years 2018-19 to 2020-
21 & valid upto 30.06.2021
Post box no. 7087, 27th cross, 12th Main, Banashankari 2nd Stage, Bengaluru- 560070, INDIA
Ph: 91-80- 26711780/81/82 Email: principal@bnmit.in, www. bnmit.org
Department of Information Science and Engineering
2020 – 2021
B.N.M. Institute of Technology
Approved by AICTE, Affiliated to VTU, Accredited as grade A Institution by NAAC.
All UG branches – CSE, ECE, EEE, ISE & Mech.E accredited by NBA for academic years 2018-19 to 2020-
21 & valid upto 30.06.2021
Post box no. 7087, 27th cross, 12th Main, Banashankari 2nd Stage, Bengaluru- 560070, INDIA
Ph: 91-80- 26711780/81/82 Email: principal@bnmit.in, www. bnmit.org
DEPARTMENT OF INFORMATION SCIENCE & ENGINEERING
CERTIFICATE
Certified that the seminar entitled “Dynamic Graph CNN for Learning on Point
Clouds” is carried out by Mr. Bharath S bearing USN 1BG17IS008 the bonafide student of
B.N.M Institute of Technology in partial fulfillment for the award of Bachelor of Engineering
in Information Science & Engineering of the Visvesvaraya Technological University,
Belagavi during the year 2020-2021. It is certified that all corrections / suggestions indicated for
Internal Assessment have been incorporated in the report. The seminar has been approved as it
satisfies the academic requirements in respect of seminar prescribed for the said Degree.
Dr. S Srividhya Dr. Shashikala
Associate Professor, Dept. of ISE Prof. & Head, Dept. of ISE
BNMIT BNMIT
ACKNOWLEDGEMENT
I consider it a privilege to express through the pages of this report, a few words of gratitude to all
those distinguished personalities who guided and inspired me in the completion of the seminar.
I would like to thank Shri. Narayan Rao R Maanay, Secretary, BNMIT, Bengaluru for providing
excellent academic environment in college.
I would like to thank Prof. T.J. Rama Murthy, Director, BNMIT, Bengaluru for having extended
his support and encouragement during work.
I would like to thank Dr. S.Y. Kulkarni, Additional Director, BNMIT, Bengaluru for having
extended his support and encouragement during work.
I would like to express my gratitude to Prof. Eishwar N Maanay, Dean Administration, for his
relentless support, guidance, and encouragement.
I would like to thank Dr. Krishnamurthy G.N., Principal, BNMIT, Bengaluru for his constant
encouragement.
I would like to thank Dr. Shashikala, Professor and Head of the department of Information
Science and Engineering, BNMIT, Bengaluru for her support and encouragement towards the
completion of seminar.
I would like to express my gratitude to my guide Dr. S Srividhya, Associate Professor,
department of Information Science and Engineering who has given me all the support and
guidance in completing the seminar work successfully.
I would like to thank seminar coordinator Mr. Manjunath G.S., Assistant Professor, department
of Information Science and Engineering, BNMIT, for being the guiding force towards successful
completion of the seminar.
Bharath S
ABSTRACT
Point clouds provide a flexible geometric representation suitable for countless applications in
computer graphics; they also comprise the raw output of most 3D data acquisition devices.
While hand-designed features on point clouds have long been proposed in graphics and vision,
however, the recent overwhelming success of convolutional neural networks (CNNs) for image
analysis suggests the value of adapting insight from CNN to the point cloud world. Point clouds
inherently lack topological information so designing a model to recover topology can enrich
the representation power of point clouds. To this end, the team propose a new neural network
module dubbed EdgeConv suitable for CNN-based high-level tasks on point clouds including
classification and segmentation. EdgeConv acts on graphs dynamically computed in each layer
of the network. It is differentiable and can be plugged into existing architectures. Compared to
existing modules operating in extrinsic space or treating each point independently, EdgeConv
has several appealing properties: It incorporates local neighbourhood information; it can be
stacked applied to learn global shape properties; and in multi-layer systems affinity in feature
space captures semantic characteristics over potentially long distances in the original
embedding. The performance of the model is evaluated on standard benchmarks including
ModelNet40, ShapeNetPart, and S3DIS.
TABLE OF CONTENTS
Chapter No Title Page No
1 Introduction 01
2 Literature Survey 06
3 Comparative Study 11
4 Conclusion 15
5 References 16
LIST OF TABLES
Chapter No. Figure No. Description Page No.
3 3.1 Classification results on ModelNet40 11
3 3.2 Complexity, forward time, and accuracy of 12
different models
3 3.3 Part segmentation results on ShapeNet part 14
dataset
LIST OF FIGURES
Chapter No. Figure No. Description Page No.
1 1.1 Overview of the network and the EdgeConv 02
operation
1 1.2 Model architecture for classification and 03
segmentation
1 1.3 Mean and overall accuracy of proposed model 04
1 1.4 Segmentation result of the proposed model with 04
PointNet result
3 3.1 Part segmentation testing results for tables, chairs 13
and lamps
CHAPTER 1
INTRODUCTION
1.1 Introduction
Point cloud is a scattered collection pf points in 3D graph, which provides a flexible
geometric representation of objects and environment suitable for countless applications in
computer graphics.
Over these long years convolutional neural network has come a long away, its recent
overwhelming success in image analysis, suggest the value of adapting CNN into analysing
point cloud, which has led the team to DGCNN.
The proposed approach is a neural network module called dubbed EdgeCov which
is suitable for CNN-based high-level tasks on point clouds including classification and
segmentation. EdgeConv acts on graphs dynamically computed in each layer of the
network. It is differentiable and can be plugged into existing architectures. The approach
proposed has several appealing properties, it incorporates local neighborhood information,
the approach can be stacked applied to learn global shape properties and in multi-layer
systems affinity in feature space captures sematic characteristics over potentially long
distances in the original embedding.
This approach was pioneered by PointNet which achieves permutation invariance
of points by operating on each point independently and subsequently applying a symmetric
function to accumulate features. Various extensions of PointNet consider neighborhoods
of points rather than acting on each independently, these allow the network to exploit local
features, improving upon performance of the basic model. These techniques largely treat
points independently at local scale to maintain permutation invariance. This independence,
however, neglects the geometric relationships among points, presenting a fundamental
limitation that cannot capture local features.
The approach proposed captures local geometric structure while maintaining
permutation invariance. Instead of generating point features directly from their
embeddingd, EdgeConv generates edge features that describes the relationships between a
point and its neighbors. EdgeConv is easy to implement and integrate into existing deep
learning models to improve their performance. In their experiments, they have integrated
EdgeConv into the basic version of PointNet without using any feature transformation.
B.E., Dept. of ISE, BNMIT Page 1 2020-21
Dynamic Graph CNN for Learning on Point Clouds
1.2 Overview of Technology
Fig 1.1: Overview of the network and the EdgeConv operation.
The approach proposed by the team s a new network model called dubbed
EdgeConv which acts on graphs dynamically computed in each layer of the network.
In the figure 1.1, to the left is the diagrammatic representation of the neural
network, where eij is the computed edge feature from pair xi and xj, hө( ) function is
instantiated using a fully connected layer, and the learnable parameters are its associated
weights.
To the right of the figure 1.1, shows the EdgeConv operation. The output of the
EdgeConv is calculated by aggregating the edge features associated with all the edge
emanating from each connected vertex. The approach proposed is inspired by PointNet and
convolution operations. Instead of working on individual points like PointNet, the approach
exploits local geometric structures by constructing a local neighbourhood graph and
applying convolution-like operations on the edges connecting neighbouring pairs of points.
The classification model takes as input n points, calculates an edge feature set of size k for
each point at an EdgeConv layer, and aggregates features within each set to compute
EdgeConv responses for corresponding points. The output features of the last EdgeConv
layer are aggregated globally to form an 1D global descriptor, which is used to generate
classification scores for c classes. The segmentation model extends the classification model
by concatenating the 1D global descriptor and all the EdgeConv outputs for each point. The
model outputs per-point classification scores for p semantic labels.
B.E., Dept. of ISE, BNMIT Page 2 2020-21
Dynamic Graph CNN for Learning on Point Clouds
Fig 1.2: Model architecture for classification and segmentation
The point cloud transform block is designed to align an input point set to a canonical space
by applying an estimated 3 × 3 matrix. To estimate the 3 × 3 matrix, a tensor concatenating
the coordinates of each point and the coordinate differences between its k neighbouring
points is used.
The EdgeConv block takes as input a tensor of shape n × f, computes edge features
for each point by applying a multi-layer perceptron (mlp) with the number of layer neurons
defined as {a1, a2, ..., an}, and generates a tensor of shape n × an after pooling among
neighbouring edge features.
Fig.1.3 shows the results for the classification task. The proposed model achieves
the best results on ModelNet40 dataset. The model provides an accuracy of 91.5% with
average class accuracy of 89.28%, the figure also shows accuracy for each individual class.
Fig.1.4 shows the results for the segmentation task. The proposed model shows a much
better segmentation compared to PointNet.
B.E., Dept. of ISE, BNMIT Page 3 2020-21
Dynamic Graph CNN for Learning on Point Clouds
Fig 1.3: Mean and overall accuracy of proposed model.
Fig 1.4: Segmentation result of the proposed model with PointNet result.
B.E., Dept. of ISE, BNMIT Page 4 2020-21
Dynamic Graph CNN for Learning on Point Clouds
1.3 Motivation
Point clouds provide a flexible geometric representation suitable for countless applications
in computer graphics; they also comprise the raw output of most 3D data acquisition
devices. While hand-designed features on point clouds have long been proposed in graphics
and vision, however, the recent overwhelming success of convolutional neural networks
(CNNs) for image analysis suggests the value of adapting insight from CNN to the point
cloud world. Point clouds inherently lack topological information so designing a model to
recover topology can enrich the representation power of point clouds. State-of-the-art deep
neural networks are designed specifically to handle the irregularity of point clouds, directly
manipulating raw point cloud data rather than passing to an intermediate regular
representation. These techniques largely treat points independently at local scale to
maintain permutation invariance. This independence, however, neglects the geometric
relationships among points, presenting a fundamental limitation that cannot capture local
features. To address these drawbacks, the team proposes a novel simple operation, called
EdgeConv, which captures local geometric structure while maintaining permutation
invariance.
1.4 Problem Statement
To propose a novel operation for learning from point clouds, to better capture local
geometric features of point clouds while still maintaining permutation invariance and
provide a model that can learn to semantically group points by dynamically updating a
graph of relationships from layer to layer.
1.5 Purpose and Scope
The purpose of the study is to explore scalable solutions to Object Detection, by using the
current state-of the-art backbone networks DGCNN. Additionally, the study proposes a
novel feature network dubbed EdgeConv. The purpose is to study the effectiveness of the
approach and this new neural network module dubbed EdgeConv suitable for CNN-based
high-level tasks on point clouds including classification and segmentation.
The scope of the study encompasses the study of the proposed architecture and feature
network, and to verify if the scaling of the same as per the dimensions laid out by DGCNN
achieve the goals without losing accuracy. The same is accomplished by performing testing
and comparing the networks result with other existing models in both classification and
segmentation.
B.E., Dept. of ISE, BNMIT Page 5 2020-21
CHAPTER 2
LITERATURE SURVEY
A literature survey in a project report represents the study done to assist in the completion
of a project. A literature survey also describes a survey of the previous existing material on
a topic of the report.
Literature surveys provide brief overviews or a summary of the current research on
topics. The structure written requires to be in a way that it seemed logical. It needs to
chronologically represent a development of the ideas in the field that is being researched.
Literature surveys are used in ensuring that the used experiments, methodologies
and experiments offer reliability and validity in the research being conducted. The surveys
need to show essential content avoiding much interpretation. One’s opinions and
conclusions require to be separated from the content in the cited sources. The topic of
literature survey must be relevant and narrow for it to be straight to the point. It identifies
the most relevant research papers from a study on the topic.
For literature survey in this report, mainly four papers are considered, they are:-
[1] PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation
[2] PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space
[3] Frustum PointNets for 3D Object Detection from RGB-D Data
[4] Comprehensive Study on LiDAR based Object Detection
These papers have made it clear and easy to understand the need of DGCNN in the field of
object detection using PCD format of data.
B.E., Dept. of ISE, BNMIT Page 6 2020-21
Dynamic Graph CNN for Learning on Point Clouds
[1] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas.. PointNet: Deep
Learning on Point Sets for 3D Classification and Segmentation. In 2017 IEEE
Conference on Computer Vision and Pattern Recognition.
PointNet has become a standard of comparison for upcoming object detection approaches
using Point Cloud Data.
• Point cloud is an important type of geometric data structure. Due to its irregular
format, most researchers transform such data to regular 3D voxel grids or
collections of images.
• Rendering PCD data unnecessarily voluminous and causes issues.
• In this paper, the team design a novel type of neural network that directly consumes
point clouds.
• The network designed respects the permutation invariance of points in the input.
• The proposed network provides a unified architecture for applications ranging from
object classification, part segmentation, to scene semantic parsing.
• Though simple, PointNet is highly efficient and effective.
• The classification network takes n points as input, applies input and feature
transformations, and then aggregates point features by max pooling.
• The segmentation network is an extension to the classification net. It concatenates
global and local features and outputs per point scores.
The key contributions of the team’s work are as follow:
• The team has designed a novel deep net architecture suitable for consuming
unordered point sets in 3D
• The team shows how their proposed net can be trained to perform 3D shape
classification, shape part segmentation and scene semantic parsing tasks.
• The team provides thorough empirical and theoretical analysis on the stability and
efficiency of their method.
• The team illustrate the 3D features computed by the selected neurons in the net and
develop intuitive explanations for its performance.
B.E., Dept. of ISE, BNMIT Page 7 2020-21
Dynamic Graph CNN for Learning on Point Clouds
[2] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas.. PointNet++: Deep
Hierarchical Feature Learning on Point Sets in a Metric Space. In AMC journal.
PointNet is a pioneer in the study of deep learning on point sets, however by design
PointNet does not capture local structures included by the metric space points, limiting its
ability to recognize fine-grained patters and generalizability to complex scenes.
• In this paper the team introduces a hierarchical neural network that applies PointNet
recursively on a nested partitioning of the input point set.
• By exploiting metric space distances, the proposed network is able to learn local
features with increasing contextual scales.
• The team proposes novel set learning layers to adaptively combine features from
multiple scales.
• Experiments show that the network called PointNet++ is able to learn deep point
set features efficiently and robustly.
• The proposed network can be viewed as an extension of PointNet with added
hierarchical structure.
• The proposed network consistes of a hierarchical structure which is composed by a
number of set abstraction levels.
• At each level in the hierarchical structure, a set of points is processed and abstracted
to produce a new set with fewer elements. The set abstraction level is made of three
key layers: Sampling layer, Grouping layer and PointNet layer.
• The Sampling layer selects a set of points from input points, which defines the
centroids of local regions.
• The Grouping layer then constructs local region sets by finding “neighbouring”
points around the centroids.
• The PointNet layer uses a mini-PointNet to encode local region patterns into feature
vectors.
• PointNet++ recursively functions on a nested partitioning of the input point set, and
is effective in learning hierarchical features with respect to the distance metric. To
handle the non-uniform point sampling issue, the team propose two novel set
abstraction layers that intelligently aggregate multi-scale information according to
local point densities. These contributions has enabled the team to achieve state-of-
the-art performance on challenging benchmarks of 3D point clouds.
B.E., Dept. of ISE, BNMIT Page 8 2020-21
Dynamic Graph CNN for Learning on Point Clouds
[3] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas.. Frustum PointNets for
3D Object Detection from RGB-D Data. In 2018 IEEE Conference on Computer Vision
and Pattern Recognition.
• In this work, the team studies 3D object detection from RGBD data in both indoor
and outdoor scenes.
• Previous methods focus on images or 3D voxels, often obscuring natural 3D
patterns and invariances of 3D data. the team directly operates on raw point clouds
by popping up RGB-D scans.
• Instead of solely relying on 3D proposals, the proposed method leverages both
mature 2D object detectors and advanced 3D deep learning for object localization,
achieving efficiency as well as high recall for even small objects.
• The method is evaluated on KITTI and SUN RGB-D 3D detection benchmarks, the
method outperforms the state of the art by remarkable margins while having real-
time capability.
• Given RGB-D data, the approach first generate 2D object region proposals in the
RGB image using a CNN. Each 2D region is then extruded to a 3D viewing frustum
in which we get a point cloud from depth data. Finally, the frustum PointNet
predicts a 3D bounding box for the object from the points in frustum.
• The proposed approach first leverage a 2D CNN object detector to propose 2D
regions and classify their content. 2D regions are then lifted to 3D and thus become
frustum proposals. Given a point cloud in a frustum, the object instance is
segmented by binary classification of each point. Based on the segmented object
point cloud (m×c), a light-weight regression PointNet (T-Net) tries to align points
by translation such that their centroid is close to amodal box center . At last the box
estimation net estimates the amodal 3D bounding box for the object.
The key contributions of the team’s work are as follow:
• The team propose a novel framework for RGB-D data-based 3D object detection
called Frustum PointNets.
• The team shows how one can train 3D object detectors under their framework and
achieve state-of-the-art performance on standard 3D object detection benchmarks.
B.E., Dept. of ISE, BNMIT Page 9 2020-21
Dynamic Graph CNN for Learning on Point Clouds
• The team provides extensive quantitative evaluations to validate their design
choices as well as rich qualitative results for understanding the strengths and
limitations of the method.
[4] Dr S Srividhya, Bharath S and Vinay S. Comprehensive Study on LiDAR based
Object Detection. In Journal of Huazhong University of Science and Technology,2021.
The computer vision is a fast paced growing technology, and there are variety of solutions
and approaches produced on a regular basis to solve real time problems. The solution with
the best accuracy is not always the best solution for a wide variety of problems and this
leads to studying these various approaches and solutions to find the best fit for a particular
problem.
• This paper gives a comprehensive view of different types of approaches taken to
detect an object using point cloud data like Dynamic Graphic Convolution Neural
Network (DGCNN), PointNet, PointNet++ and GDANet.
• The team analyses all approaches and methodologies to find a best approach to
develop a system which analyses the surrounding environment and detect the
objects for any autonomous or guided system using LiDAR technology.
• The team used ModelNet40 dataset to evaluate the different approaches and arrive
at a feasible and apropos solution to the development of the above mentioned
system.
• The 3 algorithms PointNet, PointNet++ and DGCNN, mainly concentrates on the
spatial position of the point cloud data and try to find the nearest neighbours of these
points to create a network and connect them together for classification.
• The algorithm GDANet uses a unique and different approach of disentangling the
3D points and then are flattened and contoured and fused with original points to get
a better hold on their spatial positioning in a real time scenario.
B.E., Dept. of ISE, BNMIT Page 10 2020-21
CHAPTER 3
COMPARITIVE STUDY
The comparative study chapter includes comparison of the state-of-the-art CRAFT model
with various other text detection model across different kinds of standard datasets with
specific evaluation parameters.
Here by specific evaluation parameters are:
• The dataset used for both classification and segmentation task.
• Complexity.
• Forward time.
For classification task, the team evaluates their model on the ModelNet40 dataset. The
dataset contains 12,311 meshed CAD models from 40 categories. 9,843 models are used
for training and 2,468 models are for testing.
Table 3.1. Classification results on ModelNet40
Table 3.1 shows the results for the classification task. The proposed model achieves the
best results on this dataset. By using a fixed graph the team has determined by proximity
in the input point cloud that their model is 1.0% better than PointNet++. An advanced
version including dynamical graph recomputation achieves the best results on this dataset.
B.E., Dept. of ISE, BNMIT Page 11 2020-21
Dynamic Graph CNN for Learning on Point Clouds
All the experiments are performed with point clouds that contain 1024 points except last
row. The team further tests out their model with 2048 points. The k used for 2048 points is
40 to maintain the same density. Note that PCNN uses additional augmentation techniques
like randomly sampling 1024 points out of 1200 points during both training and testing.
The team uses ModelNet40 dataset to compare the complexity of the model to previous
state-of-the-art models.
Table 3.2. Complexity, forward time, and accuracy of different models
Table 3.2 shows that the proposed model achieves the best tradeoff between the model
complexity (number of parameters), computational complexity (measured as forward pass
time), and the resulting classification accuracy.
The proposed baseline model using the fixed k-NN graph outperforms the previous state-
of-the-art PointNet++ by 1.0% accuracy, at the same time being 7 times faster. A more
advanced version of the model including a dynamically-updated graph computation
outperforms PointNet++, PCNN by 2.2% and 0.6% respectively, while being much more
efficient. The number of points in each experiment is also 1024 in this section.
For part segmentation task, the team evaluates their model on ShapeNet part dataset. For
this task, each point from a point cloud set is classified into one of a few predefined part
category labels. The dataset contains 16,881 3D shapes from 16 object categories,
annotated with 50 parts in total. 2,048 points are sampled from each training shape, and
most sampled point sets are labelled with less than six parts. The team uses Intersection-
over-Union (IoU) on points to evaluate their model and compare with other benchmarks.
The IoU of a shape is computed by averaging the IoUs of different parts occurring in that
shape, and the IoU of a category is obtained by averaging the IoUs of all the shapes
belonging to that category. The mean IoU (mIoU) is finally calculated by averaging the
IoUs of all the testing shapes.
B.E., Dept. of ISE, BNMIT Page 12 2020-21
Dynamic Graph CNN for Learning on Point Clouds
Fig 3.1. Part segmentation testing results for tables, chairs and lamps.
B.E., Dept. of ISE, BNMIT Page 13 2020-21
Dynamic Graph CNN for Learning on Point Clouds
Fig 3.1 shows visual comparison of the result of the proposed model and PointNet.
Table 3.3. Part segmentation results on ShapeNet part dataset.
Table 3.3 shows the evaluation result of part segmentation task of PointNet, PointNet++,
Kd-Net, LocalFeatureNet, PCNN and the proposed model on ShapeNet dataset. It is clearly
observed that the proposed model outperforms the pervious start-of -the-art models.
B.E., Dept. of ISE, BNMIT Page 14 2020-21
CHAPTER 4
CONCLUSION
The proposed novel operation for learning from point clouds, EdgeConv is better at
capturing local geometric features of point clouds while still maintaining permutation
invariance.
The model can learn to semantically group points by dynamically updating a graph of
relationships from layer to layer and can be integrated into multiple existing pipelines for
point cloud processing.
By presenting extensive analysis and testing of EdgeConv the team shows that it achieves
state-of-the-art performance on benchmark datasets. And for future research the team has
made sure that the code can be easily understandable and have released it with an open
licence.
B.E., Dept. of ISE, BNMIT Page 15 2020-21
CHAPTER 5
REFERENCES
[1] Wang, Y.; Sun, Y.; Liu, Z.; Sarma, S.E.; Bronstein, M.M.; Solomon, J.M. Dynamic
graph CNN for learning on point clouds. arXiv (2018), arXiv:1801.07829.
[2] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas.. PointNet: Deep
Learning on Point Sets for 3D Classification and Segmentation. In 2017 IEEE
Conference on Computer Vision and Pattern Recognition
[3] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas.. PointNet++: Deep
Hierarchical Feature Learning on Point Sets in a Metric Space. In AMC journal.
[4] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas.. Frustum PointNets for
3D Object Detection from RGB-D Data. In 2018 IEEE Conference on Computer Vision
and Pattern Recognition.
[5] Dr S Srividhya, Bharath S and Vinay S. Comprehensive Study on LiDAR based
Object Detection. In Journal of Huazhong University of Science and Technology,2021.
B.E., Dept. of ISE, BNMIT Page 16 2020-21
