This document discusses object detection using deep neural networks. It describes different types of neural networks including convolutional neural networks (CNNs), which are well-suited for object detection tasks. CNNs use techniques like parameter sharing and sparse connections to recognize visual patterns. Popular object detection algorithms that use CNNs are R-CNN, which proposes regions and classifies them, and YOLO (You Only Look Once), which frames detection as a single regression problem from image to bounding boxes. While very fast, YOLO struggles with small or grouped objects and unusual aspect ratios.

1. 1
OBJECT DETECTION USING DEEP
NEURAL NETWORK

2. 2
What is deep neural networks?

A technology built to simulate the activity of the
human brain.

Has several layers including input & output layers.

Each layer performs specific types of functions.

Uses in image recognition, verification,
classification, object detection, real-time object
detection, face recognition etc.

3. 3
TYPES OF NEURAL NETWORKS

Artificial Neural Network (ANN)

Deep Residual Network (Deep ResNet)

Recurrent Neural Network (RNN)

Convolutional Neural Network (CNN)

4. 4
CONVOLUTIONAL NEURAL NETWORK(CNN)

One of the most popular neural networks.

Most commonly used in images recognition,
images classifications. Object detections,
recognition faces, computer vision etc.

Types of CNN:

LeNet-5

AlexNet

VGG-16

Inception Network

GoogleNet etc.

5. 5
CNN IN OBJECT DETECTION

Very good for recognizing patterns such as edges
(vertical/horizontal), shapes, colours, and textures.

Most popular neural network for object detection.

Why CNN?:
 Parameter sharing
 Sparsity of connections

6. 6
OBJECT DETECTION ALGORITHMS
Some popular object detection algorithms are:

Deformable parts models(DPM)

R-CNN(Region-CNN)

YOLO(You Only Look Once)

7. 7
Deformable parts models(DPM)

Use a sliding window approach to object detection

DPM uses a disjoint pipeline to -
 extract static features
 classify regions
 predict bounding boxes for high scoring regions, etc.

8. 8
Fig: Deformable parts models

9. 9
R-CNN(Region-CNN)

One of the most popular object detection algorithm

Uses region proposals in-stead of sliding windows

First generate potential bounding boxes in an
image

Run a classifier on these proposed boxes.

Post-processing is used to refine the bounding
boxes, eliminate duplicate detections, and rescore
the boxes based on other objects in the scene.

10. 10

11. 11
LIMITATIONS OF R-CNN

Complex pipelines are slow

Hard to optimize

Individual component must be trained separately

12. 12
YOLO(YOU ONLY LOOK ONCE)
According to the paper “You Only Look Once:Unified,
Real-Time Object Detection” - by Joseph Redmon,
Santosh Divvala, Ross Girshick, Ali Farhadi:

Reframe object detection as a single regression
problem

Straight from image pixels to bounding box
coordinates and class probabilities

13. 13
METHODOLOGY

A single convNet simultaneously predicts multiple
bounding boxes

Calculate class probabilities for those boxes

Divides the input image into an S × S grid.

center of an object is responsible for detecting
that object.

Each grid cell predicts B bounding boxes and
confidence scores for those boxes.

Confidence scores reflect how confident the model
is that the box contains an object

14. 14
METHODOLOGY
Fig:YOLO object detection

15. 15
METHODOLOGY

How accurate it thinks the box is that it predicts

Confidence can be defined as Pr(Object) ∗ IOU
truth

Each bounding box consists of 5 predictions: x, y,
w, h,and confidence.

16. 16
Fig: YOLO convolution layers

17. 17
WHY YOLO?

YOLO is extremely fast

It doesn’t need a complex pipeline

YOLO achieves more than twice the mean average
precision of other real-time systems.

YOLO sees the entire image during training and
test time

YOLO makes less than half the number of
background errors compared to Fast R-CNN.

YOLO outperforms top detection methods like DPM
and R-CNN by a wide margin

18. 18
LIMITATIONS

YOLO imposes strong spatial constraints on
bounding box predictions.

Model struggles with small objects that appear in
groups, such as flocks of birds.

It struggles to generalize to objects in new or
unusual aspect ratios or configurations.

Main source of error is incorrect localizations.