Intro to AI Platform

Август 22, 2022

Содержание

2. Contents SSD(Single Shot Detector) ? SSD Implementation TensorFlow Hub & TF Object Detection API ? Object
3. SSD SSD (Single Shot Detector) : needs only one single shot to detect multiple objects within
4. SSD : use multiscale feature maps (38x38, 19x19, 10x10, 5x5, 3x3, 1x1) : feature maps at
5. SSD Implementation Feature map size : 19x19x1024 ? 1x1 conv ? 19x19x256 ? 3x3 conv with
6. SSD Implementation Feature maps Apply 3x3 conv w/ stride 1, padding 1 ? Concat ? prediction
7. TensorFlow Hub : In the last lab lecture, we use pre-trained ResNet50 which is provided by
8. Object Detection Inference Tutorial 1. Import tensorflow_hub with other necessary libraries 2. Define some other additional
9. Object Detection Inference Tutorial 3. Define image_url of pictures what you want to know which objects
10. Object Detection Inference Tutorial 5. Define functions for loading image and performing detection & Run detection
11. Object Detection Inference Tutorial 6. You can perform inference on some additional images Please try with
12. Inference Tutorial 2 1. install the Object Detection API the part for importing and defining a
13. Inference Tutorial 2 4. Define model what we want to use & load it from TensorFlow
14. Inference Tutorial 2 5. select image, or you can edit image_path for detecting your own images
15. Inference Tutorial 2 6. Do the Inference & Visualize the result
17. Скачать презентацию

Слайд 2

Contents
SSD(Single Shot Detector) ?
SSD Implementation
TensorFlow Hub & TF Object Detection API

?
Object Detection Inference Tutorial (1), (2)

Слайд 3

SSD
SSD (Single Shot Detector)
: needs only one single shot to detect

multiple objects within the image (not uses region proposal)
: composes of 2 parts ? extract feature maps + apply convolution filters for detection
: backbone model + SSD head
: backbone model usually is a pre-trained image classification network as a feature extractor
(e.g., VGG, ResNet, MobileNet, etc.)

Слайд 4

SSD
: use multiscale feature maps
(38x38, 19x19, 10x10, 5x5, 3x3, 1x1)
: feature

maps at the beginning ? small objects
: feature maps at the end ? large objects
? Objects of different sizes can be detected through different feature maps

5x5 feature map

3x3 feature map

Слайд 5

SSD Implementation
Feature map size : 19x19x1024
? 1x1 conv ? 19x19x256
? 3x3

conv with stride 2 ? 10x10x512

Class SSD

Слайд 6

SSD Implementation
Feature maps
Apply 3x3 conv w/ stride 1, padding 1
? Concat

? prediction

Final Output

Слайд 7

TensorFlow Hub
: In the last lab lecture, we use pre-trained ResNet50

which is provided by tf.keras.applications
: let’s try to use TensorFlow Hub for Object detection
TensorFlow Hub
: a repository of “trained machine learning models”
? We can load model from TF Hub and use it for fine-tuning or inference
( https://tfhub.dev )
TensorFlow Object Detection API(Application Programming Interface)
: an open source framework built on top of TensorFlow that makes it easy to construct, train and deploy object detection models
: provides useful functions or classes for object detection

Слайд 8

Object Detection Inference Tutorial
1. Import tensorflow_hub with other necessary libraries
2. Define

some other additional functions

Слайд 9

Object Detection Inference Tutorial
3. Define image_url of pictures what you want

to know which objects are in it

4. Load model from TF Hub

Of course, other models in the hub are available, but let’s apply ‘SSD’ we learned

Simple way to load pre-trained models

Слайд 10

Object Detection Inference Tutorial
5. Define functions for loading image and performing

detection & Run detection

Слайд 11

Object Detection Inference Tutorial
6. You can perform inference on some additional

images

Please try with your own images
- if you have URL of image, edit this cell
- Or, drag your images into proper directory in Colab like in Lab2, and edit ‘downloaded_image_path’

Слайд 12