MML

Image Segmentation and

Detection
Executive Program on AI/ML
IIIT Hyderabad

TECHNOLOGY INNOVATION HUB ON DATA DRIVEN

 Computer Vision Tasks
Semantic Object Instance
Classification
Segmentation Detection Segmentation

CAT GRASS, CAT, DOG, DOG, CAT DOG, DOG, CAT

TREE, SKY

No spatial extent No objects, just pixels Multiple Object This image is CC0 public domain

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 2

 Semantic Segmentation
Semantic Object Instance
Classification
Segmentation Detection Segmentation

CAT GRASS, CAT, DOG, DOG, CAT DOG, DOG, CAT

TREE, SKY

No spatial extent No objects, just pixels Multiple Object

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 3

 Semantic Segmentation: The Problem

GRASS, CAT, At test time, classify each pixel of a new image.

TREE, SKY, ...
Paired training data: for each training image,
each pixel is labeled with a semantic category.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 4

 Semantic Segmentation Idea: Sliding Window

Full image

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 5

 Semantic Segmentation Idea: Sliding Window

Full image

Impossible to classify without context

Q: how do we include context?

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 6

 Semantic Segmentation Idea: Sliding Window

Full image

Q: how do we model this?

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 7

 Semantic Segmentation Idea: Sliding Window
Classify center
Extract patch pixel with CNN

Full image
Cow

Cow

Grass

Farabet et al, “Learning Hierarchical Features for Scene Labeling,” TPAMI 2013
Pinheiro and Collobert, “Recurrent Convolutional Neural Networks for Scene Labeling”, ICML 2014

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 8

 Semantic Segmentation Idea: Sliding Window
Classify center
Extract patch pixel with CNN

Full image
Cow

Cow

Grass
Problem: Very inefficient! Not
reusing shared features between
overlapping patches Farabet et al, “Learning Hierarchical Features for Scene Labeling,” TPAMI 2013
Pinheiro and Collobert, “Recurrent Convolutional Neural Networks for Scene Labeling”, ICML 2014

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 10

 Semantic Segmentation Idea: Convolution

Full image

An intuitive idea: encode the entire image with conv net, and do semantic segmentation
on top.

Problem: classification architectures often reduce feature spatial sizes to go deeper, but
semantic segmentation requires the output size to be the same as input size.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 10

 Semantic Segmentation Idea: Fully Convolutional
Design a network with only convolutional layers
without downsampling operators to make predictions
for pixels all at once!

Conv Conv Conv Conv argmax

Input:
Scores: Predictions:
3xHxW
CxHxW HxW
Convolutions:
DxHxW

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 11

 Semantic Segmentation Idea: Fully Convolutional
Design a network with only convolutional layers
without downsampling operators to make predictions
for pixels all at once!

Conv Conv Conv Conv argmax

Input:
Scores: Predictions:
3xHxW
CxHxW HxW
Convolutions:
Problem: convolutions at DxHxW
original image resolution will
be very expensive ...

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 12

 Semantic Segmentation Idea: Fully Convolutional
Design network as a bunch of convolutional layers, with
downsampling and upsampling inside the network!

Med-res: Med-res:
D2 x H/4 x W/4 D2 x H/4 x W/4

Low-res:
D3 x H/4 x W/4
Input: High-res: High-res: C x H x W Predictions:
3xHxW D1 x H/2 x W/2 D1 x H/2 x W/2 HxW

Long, Shelhamer, and Darrell, “Fully Convolutional Networks for Semantic Segmentation”, CVPR 2015
Noh et al, “Learning Deconvolution Network for Semantic Segmentation”, ICCV 2015

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 13

 Semantic Segmentation Idea: Fully Convolutional
Downsampling: Design network as a bunch of convolutional layers, with Upsampling:
Pooling, strided downsampling and upsampling inside the network! ???
convolution
Med-res: Med-res:
D2 x H/4 x W/4 D2 x H/4 x W/4

Low-res:
D3 x H/4 x W/4
Input: High-res: High-res: C x H x W Predictions:
3xHxW D1 x H/2 x W/2 D1 x H/2 x W/2 HxW

Long, Shelhamer, and Darrell, “Fully Convolutional Networks for Semantic Segmentation”, CVPR 2015
Noh et al, “Learning Deconvolution Network for Semantic Segmentation”, ICCV 2015

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 14

 In-Network upsampling: “Unpooling”

Nearest Neighbor “Bed of Nails”

1 1 2 2 1 0 2 0

1 2 1 1 2 2 1 2 0 0 0 0

3 4 3 3 4 4 3 4 3 0 4 0

3 3 4 4 0 0 0 0

Input: 2 x 2 Output: 4 x 4 Input: 2 x 2 Output: 4 x 4

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 15

 In-Network upsampling: “Max Unpooling”
Max Pooling Max Unpooling
Remember which element was max! Use positions from
1 2 6 3 pooling layer 0 0 2 0

3 5 2 1 5 6
… 1 2 0 1 0 0

1 2 2 1 7 8 3 4 0 0 0 0
Rest of the network
7 3 4 8 3 0 0 4

Input: 4 x 4 Output: 2 x 2 Input: 2 x 2 Output: 4 x 4

Corresponding pairs of
downsampling and
upsampling layers

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 16

 Learnable Upsampling
Recall: Normal 3 x 3 convolution, stride 1 pad 1

Input: 4 x 4 Output: 4 x 4

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 17

 Learnable Upsampling
Recall: Normal 3 x 3 convolution, stride 1 pad 1

Dot product
between filter
and input

Input: 4 x 4 Output: 4 x 4

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 18

 Learnable Upsampling
Recall: Normal 3 x 3 convolution, stride 1 pad 1

Dot product
between filter
and input

Input: 4 x 4 Output: 4 x 4

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 19

 Learnable Upsampling
Recall: Normal 3 x 3 convolution, stride 2 pad 1

Input: 4 x 4 Output: 2 x 2

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 20

 Learnable Upsampling
Recall: Normal 3 x 3 convolution, stride 2 pad 1

Dot product
between filter
and input

Input: 4 x 4 Output: 2 x 2

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 21

 Learnable Upsampling
• Recall: Normal 3 x 3 convolution, stride 2 pad 1

Filter moves 2 pixels in

the input for every one
pixel in the output.
Dot product
between filter Stride gives ratio between
and input movement in input and
output

We can interpret strided

Input: 4 x 4 Output: 2 x 2 convolution as “learnable
downsampling”.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 22

 Learnable Upsampling: Transposed Convolution
3 x 3 transposed convolution, stride 2 pad 1

Input: 2 x 2 Output: 4 x 4

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 23

 Learnable Upsampling: Transposed Convolution
3 x 3 transposed convolution, stride 2 pad 1

Input gives
weight for
filter

Input: 2 x 2 Output: 4 x 4

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 24

 Learnable Upsampling: Transposed Convolution
3 x 3 transposed convolution, stride 2 pad 1

Filter moves 2 pixels in

Input gives the output for every one
weight for pixel in the input
filter
Stride gives ratio between
movement in output and
input
Input: 2 x 2 Output: 4 x 4

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 25

 Learnable Upsampling: Transposed Convolution
Sum where
3 x 3 transposed convolution, stride 2 pad 1 output overlaps

Filter moves 2 pixels in

Input gives the output for every one
weight for pixel in the input
filter
Stride gives ratio between
movement in output and
input
Input: 2 x 2 Output: 4 x 4

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 26

 Semantic Segmentation Idea: Fully Convolutional
Upsampling:
Downsampling: Design network as a bunch of convolutional layers, with
Unpooling or strided
Pooling, strided downsampling and upsampling inside the network!
transposed convolution
convolution
Med-res: Med-res:
D2 x H/4 x W/4 D2 x H/4 x W/4

Low-res:
D3 x H/4 x W/4
Input: High-res: High-res: Predictions:
3xHxW D1 x H/2 x W/2 D1 x H/2 x W/2 HxW

Long, Shelhamer, and Darrell, “Fully Convolutional Networks for Semantic Segmentation”, CVPR 2015
Noh et al, “Learning Deconvolution Network for Semantic Segmentation”, ICCV 2015

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 27

 Semantic Segmentation: Summary

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 28

U-Net
 Semantic Segmentation This image is CC0 public domain

Label each pixel in the

image with a category
label

s
Sky Sky

ee
Tr

Tr
ee
s
Don’t differentiate
Cat Cow
instances, only care about
pixels
Grass Grass

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 30

 Object Detection
Semantic Object Instance
Classification
Segmentation Detection Segmentation

CAT GRASS, CAT, DOG, DOG, CAT DOG, DOG, CAT

TREE, SKY

No spatial extent No objects, just pixels Multiple Object

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 31

 Object Detection
Semantic Object Instance
Classification
Segmentation Detection Segmentation

CAT GRASS, CAT, DOG, DOG, CAT DOG, DOG, CAT

TREE, SKY

No spatial extent No objects, just pixels Multiple Object

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 32

 Object Detection: Single Object
(Classification + Localization)
Class Scores
Fully Cat: 0.9
Connected: Dog: 0.05
4096 to 1000 Car: 0.01
x, y ...

w
Vector: Fully
This image is CC0 public domain
Connected:
4096 4096 to 4 Box
Coordinates
(x, y, w, h)

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 33

 Object Detection: Single Object Correct label:

(Classification + Localization) Cat

Class Scores
Fully Cat: 0.9 Softmax
Connected: Dog: 0.05 Loss
4096 to 1000 Car: 0.01
x, y ...

w
Vector: Fully
This image is CC0 public domain
Connected:
4096 4096 to 4 Box
Coordinates L2 Loss
(x, y, w, h)
Treat localization as a
regression problem! Correct box:
(x’, y’, w’, h’)

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 34

 Object Detection: Single Object Correct label:

(Classification + Localization) Cat

Class Scores
Fully Cat: 0.9 Softmax
Connected: Dog: 0.05 Loss
4096 to 1000 Car: 0.01
x, y ...

h Multitask Loss + Loss

w
Vector: Fully
This image is CC0 public domain
Connected:
4096 4096 to 4 Box
Coordinates L2 Loss
(x, y, w, h)
Treat localization as a
regression problem! Correct box:
(x’, y’, w’, h’)

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 35

 Object Detection: Multiple Objects
CAT: (x, y, w, h)

DOG: (x, y, w, h)
DOG: (x, y, w, h)
CAT: (x, y, w, h)

DUCK: (x, y, w, h)
DUCK: (x, y, w, h)
….

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 36

 Object Detection: Multiple Objects Each image needs a different number of
outputs!

CAT: (x, y, w, h) 4 numbers

DOG: (x, y, w, h)
DOG: (x, y, w, h) 12 numbers
CAT: (x, y, w, h)

DUCK: (x, y, w, h) Many

DUCK: (x, y, w, h) numbers!
….

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 37

 Object Detection: Multiple Objects
Apply a CNN to many different crops of the
image, CNN classifies each crop as object
or background

Dog? NO
Cat? NO
Background? YES

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 38

 Object Detection: Multiple Objects
Apply a CNN to many different crops of the
image, CNN classifies each crop as object
or background

Dog? YES
Cat? NO
Background? NO

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 39

 Object Detection: Multiple Objects
Apply a CNN to many different crops of the
image, CNN classifies each crop as object
or background

Dog? YES
Cat? NO
Background? NO

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 40

 Object Detection: Multiple Objects
Apply a CNN to many different crops of the
image, CNN classifies each crop as object
or background

Dog? NO
Cat? YES
Background? NO

Q: What’s the problem with this approach?

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 41

 Object Detection: Multiple Objects
Apply a CNN to many different crops of the
image, CNN classifies each crop as object
or background

Dog? NO
Cat? YES
Background? NO

Problem: Need to apply CNN to huge

number of locations, scales, and aspect
ratios, very computationally expensive!

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 42

 Region Proposals: Selective Search
● Find “blobby” image regions that are likely to contain objects
● Relatively fast to run; e.g. Selective Search gives 2000 region
proposals in a few seconds on CPU

Alexe et al, “Measuring the objectness of image windows”, TPAMI 2012

Uijlings et al, “Selective Search for Object Recognition”, IJCV 2013
Cheng et al, “BING: Binarized normed gradients for objectness estimation at 300fps”, CVPR 2014
Zitnick and Dollar, “Edge boxes: Locating object proposals from edges”, ECCV 2014

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 43

 R-CNN

Girshick et al, “Rich feature hierarchies for accurate object detection and
semantic segmentation”, CVPR 2014.
Input image Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 44

 R-CNN

Regions of Interest
(RoI) from a proposal
method (~2k)
Girshick et al, “Rich feature hierarchies for accurate object detection and
semantic segmentation”, CVPR 2014.
Input image Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - April 26,45

2022
 R-CNN

Warped image regions

(224x224 pixels)
Regions of Interest
(RoI) from a proposal
method (~2k)
Girshick et al, “Rich feature hierarchies for accurate object detection and
semantic segmentation”, CVPR 2014.
Input image Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 46

 R-CNN

ConvN Forward each region

et through ConvNet
ConvN
(ImageNet-pretranied)
et
ConvN
et Warped image regions
(224x224 pixels)
Regions of Interest
(RoI) from a proposal
method (~2k)
Girshick et al, “Rich feature hierarchies for accurate object detection and
semantic segmentation”, CVPR 2014.
Input image Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 47

 R-CNN
SVMs Classify regions with
SVMs SVMs

SVMs
ConvN Forward each region
et through ConvNet
ConvN
(ImageNet-pretranied)
et
ConvN
et Warped image regions
(224x224 pixels)
Regions of Interest
(RoI) from a proposal
method (~2k)
Girshick et al, “Rich feature hierarchies for accurate object detection and
semantic segmentation”, CVPR 2014.
Input image Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 48

 Predict “corrections” to the RoI: 4 numbers: (dx, dy, dw, dh)
R-CNN
Bbox reg SVMs Classify regions with
Bbox reg SVMs SVMs

Bbox reg SVMs

ConvN Forward each region
et through ConvNet
ConvN
(ImageNet-pretranied)
et
ConvN
et Warped image regions
(224x224 pixels)
Regions of Interest
(RoI) from a proposal
method (~2k)
Girshick et al, “Rich feature hierarchies for accurate object detection and
semantic segmentation”, CVPR 2014.
Input image Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 49

 Predict “corrections” to the RoI: 4 numbers: (dx, dy, dw, dh)
R-CNN
Bbox reg SVMs Classify regions with
Problem: Very slow!
Bbox reg SVMs SVMs
Need to do ~2k
Bbox reg SVMs independent forward
ConvN Forward each
passes for each image!
et region through
ConvN
ConvNet
et
ConvN
et Warped image regions
(224x224 pixels)
Regions of Interest
(RoI) from a proposal
method (~2k)
Girshick et al, “Rich feature hierarchies for accurate object detection and
semantic segmentation”, CVPR 2014.
Input image Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 50

 Predict “corrections” to the RoI: 4 numbers: (dx, dy, dw, dh)
“Slow” R-CNN
Bbox reg SVMs Classify regions with
Problem: Very slow!
Bbox reg SVMs SVMs
Need to do ~2k
Bbox reg SVMs independent forward
ConvN Forward each
passes for each image!
et region through
ConvN
ConvNet
et Idea: Pass the
ConvN
image through
et Warped image regions
convnet before
(224x224 pixels)
cropping! Crop the
Regions of Interest conv feature instead!
(RoI) from a proposal
method (~2k)
Girshick et al, “Rich feature hierarchies for accurate object detection and
semantic segmentation”, CVPR 2014.
Input image Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 51

 Fast R-CNN

“Slow” R-CNN

Input image

Girshick, “Fast R-CNN”, ICCV 2015. Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 52

 Fast R-CNN

“Slow” R-CNN

“conv5” features

Run whole image

“Backbone” through ConvNet
network:
AlexNet, VGG, ConvNet
ResNet, etc Input image

Girshick, “Fast R-CNN”, ICCV 2015. Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 53

 Fast R-CNN

Regions of “Slow” R-CNN

Interest (RoIs)
from a proposal
method “conv5” features

Run whole image

“Backbone” through ConvNet
network:
AlexNet, VGG, ConvNet
ResNet, etc Input image

Girshick, “Fast R-CNN”, ICCV 2015. Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 54

 Fast R-CNN

Regions of “Slow” R-CNN

Interest (RoIs)
Crop + Resize features
from a proposal
method “conv5” features

Run whole image

“Backbone” through ConvNet
network:
AlexNet, VGG, ConvNet
ResNet, etc Input image

Girshick, “Fast R-CNN”, ICCV 2015. Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 55

 Fast R-CNN
Object Linear +
softmax Linear Box offset
category
Regions of CNN Per-Region Network “Slow” R-CNN
Interest (RoIs)
Crop + Resize features
from a proposal
method “conv5” features

Run whole image

“Backbone” through ConvNet
network:
AlexNet, VGG, ConvNet
ResNet, etc Input image

Girshick, “Fast R-CNN”, ICCV 2015. Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 56

 Fast R-CNN
Object Linear +
softmax Linear Box offset
category
Regions of CNN Per-Region Network “Slow” R-CNN
Interest (RoIs)
Crop + Resize features
from a proposal
method “conv5” features

Run whole image

“Backbone” through ConvNet
network:
AlexNet, VGG, ConvNet
ResNet, etc Input image

Girshick, “Fast R-CNN”, ICCV 2015. Figure copyright Ross Girshick, 2015; source. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 57

 Cropping Features: RoI Pool

CNN

Input Image Image features: C x H x W

(e.g. 3 x 640 x 480) (e.g. 512 x 20 x 15)

Girshick, “Fast R-CNN”, ICCV 2015.

Girshick, “Fast R-CNN”, ICCV 2015.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 58

 Cropping Features: RoI Pool
Project proposal
onto features

CNN

Input Image Image features: C x H x W

(e.g. 3 x 640 x 480) (e.g. 512 x 20 x 15)

Girshick, “Fast R-CNN”, ICCV 2015.

Girshick, “Fast R-CNN”, ICCV 2015.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 59

 Cropping Features: RoI Pool “Snap” to
grid cells
Project proposal
onto features

CNN

Input Image Image features: C x H x W

(e.g. 3 x 640 x 480) (e.g. 512 x 20 x 15)

Girshick, “Fast R-CNN”, ICCV 2015.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 60

 Cropping Features: RoI Pool “Snap” to
grid cells
Project proposal
onto features

Q: how do we resize the 512

x 5 x 4 region to, e.g., a 512
x 2 x 2 tensor?.
CNN

Input Image Image features: C x H x W

(e.g. 3 x 640 x 480) (e.g. 512 x 20 x 15)

Girshick, “Fast R-CNN”, ICCV 2015.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 61

 Cropping Features: RoI Pool “Snap” to
Divide into 2x2
grid of (roughly)
grid cells equal subregions
Project proposal
onto features

Q: how do we resize the 512

x 5 x 4 region to, e.g., a 512
x 2 x 2 tensor?.
CNN

Input Image Image features: C x H x W

(e.g. 3 x 640 x 480) (e.g. 512 x 20 x 15)

Girshick, “Fast R-CNN”, ICCV 2015.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 62

 Cropping Features: RoI Pool “Snap” to
Divide into 2x2
grid of (roughly)
grid cells equal subregions
Project proposal
onto features
Max-pool within
each subregion

CNN

Region features
(here 512 x 2 x 2;
In practice e.g 512 x 7 x 7)
Input Image Image features: C x H x W Region features always the
(e.g. 3 x 640 x 480) (e.g. 512 x 20 x 15) same size even if input
regions have different sizes!
Girshick, “Fast R-CNN”, ICCV 2015.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 63

 Cropping Features: RoI Pool “Snap” to
Divide into 2x2
grid of (roughly)
grid cells equal subregions
Project proposal
onto features
Max-pool within
each subregion

CNN

Region features
(here 512 x 2 x 2;
In practice e.g 512 x 7 x 7)
Input Image Image features: C x H x W Region features always the
(e.g. 3 x 640 x 480) (e.g. 512 x 20 x 15) same size even if input
regions have different sizes!
Girshick, “Fast R-CNN”, ICCV 2015.
Problem: Region features slightly misaligned

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 68

 Cropping Features: RoI Align
No “snapping”!
Project proposal
onto features

CNN

Input Image Image features: C x H x W

(e.g. 3 x 640 x 480) (e.g. 512 x 20 x 15)

He et al, “Mask R-CNN”, ICCV 2017

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 65

 Sample at regular points in
Cropping Features: RoI Align each subregion using
No “snapping”! bilinear interpolation
Project proposal
onto features

CNN

Input Image Image features: C x H x W

(e.g. 3 x 640 x 480) (e.g. 512 x 20 x 15)

He et al, “Mask R-CNN”, ICCV 2017

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 66

 Sample at regular points in
Cropping Features: RoI Align each subregion using
No “snapping”! bilinear interpolation
Project proposal
onto features

CNN

Feature fxy for point (x, y)

is a linear combination of
Input Image Image features: C x H x W features at its four
(e.g. 3 x 640 x 480) (e.g. 512 x 20 x 15) neighboring grid cells:

He et al, “Mask R-CNN”, ICCV 2017

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 67

 Sample at regular points in
Cropping Features: RoI Align each subregion using
No “snapping”! bilinear interpolation
Project proposal
onto features
Max-pool within
each subregion

CNN

Region features
(here 512 x 2 x 2;
In practice e.g 512 x 7 x 7)
Input Image Image features: C x H x W
(e.g. 3 x 640 x 480) (e.g. 512 x 20 x 15)

He et al, “Mask R-CNN”, ICCV 2017

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 68

 R-CNN vs Fast R-CNN

Girshick et al, “Rich feature hierarchies for accurate object detection and semantic segmentation”, CVPR 2014.
He et al, “Spatial pyramid pooling in deep convolutional networks for visual recognition”, ECCV 2014
Girshick, “Fast R-CNN”, ICCV 2015

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 69

 R-CNN vs Fast R-CNN

Problem:
Runtime dominated
by region proposals!

Girshick et al, “Rich feature hierarchies for accurate object detection and semantic segmentation”, CVPR 2014.
He et al, “Spatial pyramid pooling in deep convolutional networks for visual recognition”, ECCV 2014
Girshick, “Fast R-CNN”, ICCV 2015

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 70

 Faster R-CNN:
Make CNN do proposals!

Insert Region Proposal

Network (RPN) to predict
proposals from features

Otherwise same as Fast R-CNN:

Crop features for each proposal,
classify each one

Ren et al, “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks”, NIPS 2015
Figure copyright 2015, Ross Girshick; reproduced with permission

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 71

 Region Proposal Network

CNN

Input Image
(e.g. 3 x 640 x 480) Image features
(e.g. 512 x 20 x 15)

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 72

 Region Proposal Network Imagine an anchor box
of fixed size at each
point in the feature map

CNN

Input Image
(e.g. 3 x 640 x 480) Image features
(e.g. 512 x 20 x 15)

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 73

 Region Proposal Network Imagine an anchor box
of fixed size at each
point in the feature map

Anchor has an
object?
CNN Conv 1 x 20 x 15

At each point, predict

Input Image
(e.g. 3 x 640 x 480)
whether the corresponding
Image features
(e.g. 512 x 20 x 15) anchor contains an object
(binary classification)

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 74

 Region Proposal Network Imagine an anchor box
of fixed size at each
point in the feature map

Anchor has an
object?
CNN Conv 1 x 20 x 15
Box corrections
4 x 20 x 15

For positive boxes, also predict

Input Image
(e.g. 3 x 640 x 480)
a corrections from the anchor to
Image features
(e.g. 512 x 20 x 15) the ground-truth box (regress 4
numbers per pixel)

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 75

 In practice use K different
Region Proposal Network anchor boxes of different
size / scale at each point

Anchor has an
object?
CNN Conv K x 20 x 15
Box transforms
4K x 20 x 15

Input Image
(e.g. 3 x 640 x 480) Image features
(e.g. 512 x 20 x 15)

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 76

 In practice use K different
Region Proposal Network anchor boxes of different
size / scale at each point

Anchor has an
object?
CNN Conv K x 20 x 15
Box transforms
4K x 20 x 15

Sort the K*20*15 boxes by

Input Image their “objectness” score,
(e.g. 3 x 640 x 480) Image features take top ~300 as our
(e.g. 512 x 20 x 15) proposals

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 77

 Faster R-CNN:
Make CNN do proposals!

Jointly train with 4 losses:

1. RPN classify object / not object
2. RPN regress box coordinates
3. Final classification score (object
classes)
4. Final box coordinates

Ren et al, “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks”, NIPS 2015
Figure copyright 2015, Ross Girshick; reproduced with permission

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 78

 Faster R-CNN:
Make CNN do proposals!

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 79

 Faster R-CNN:
Make CNN do
proposals!
Glossing over many details:
- Ignore overlapping proposals with
non-max suppression
- How are anchors determined?
- How do we sample positive /
negative samples for training the
RPN?
- How to parameterize bounding
box regression?

Ren et al, “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks”, NIPS 2015
Figure copyright 2015, Ross Girshick; reproduced with permission

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 80

 Faster R-CNN:
Make CNN do
proposals!
Faster R-CNN is a
Two-stage object detector

First stage: Run once per image

- Backbone network
- Region proposal network

Second stage: Run once per region

- Crop features: RoI pool / align
- Predict object class
- Prediction bbox offset

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 81

 Faster R-CNN:
Make CNN do Do we really need
the second stage?
proposals!
Faster R-CNN is a
Two-stage object detector

First stage: Run once per image

- Backbone network
- Region proposal network

Second stage: Run once per region

- Crop features: RoI pool / align
- Predict object class
- Prediction bbox offset

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 82

 Single-Stage Object Detectors: YOLO / SSD / RetinaNet
Within each grid cell:
- Regress from each of the B
base boxes to a final box
with 5 numbers:
(dx, dy, dh, dw, confidence)
- Predict scores for each of C
classes (including
background as a class)
- Looks a lot like RPN, but
category-specific!
Input image Divide image into grid
3xHxW 7x7
Imagine a set of base
Output:
Redmon et al, “You Only Look Once:
Unified, Real-Time Object Detection”, CVPR 2016 boxes centered at each 7 x 7 x (5 * B + C)
Liu et al, “SSD: Single-Shot MultiBox Detector”, ECCV 2016
Lin et al, “Focal Loss for Dense Object Detection”, ICCV 2017 grid cell Here B = 3

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 83

 Single-Stage Object Detectors: YOLO / SSD / RetinaNet
Within each grid cell:
- Regress from each of the B
base boxes to a final box
with 5 numbers:
(dx, dy, dh, dw, confidence)
- Predict scores for each of C
classes (including
background as a class)
- Looks a lot like RPN, but
category-specific!
YOLO
Redmon et al, “You Only Look Once:
Unified, Real-Time Object Detection”, CVPR 2016
Liu et al, “SSD: Single-Shot MultiBox Detector”, ECCV 2016
Lin et al, “Focal Loss for Dense Object Detection”, ICCV 2017

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 84

 Object Detection: Lots of variables ...
Backbone “Meta-Architecture” Takeaways
Network Two-stage: Faster R-CNN Faster R-CNN is slower
VGG16 Single-stage: YOLO / SSD but more accurate
ResNet-101 Hybrid: R-FCN
Inception V2 SSD is much faster but
Inception V3 Image Size not as accurate
Inception # Region Proposals
ResNet … Bigger / Deeper
MobileNet backbones work better
Huang et al, “Speed/accuracy trade-offs for modern convolutional object detectors”, CVPR 2017

R-FCN: Dai et al, “R-FCN: Object Detection via Region-based Fully Convolutional Networks”, NIPS 2016
Inception-V2: Ioffe and Szegedy, “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift”, ICML 2015
Inception V3: Szegedy et al, “Rethinking the Inception Architecture for Computer Vision”, arXiv 2016
Inception ResNet: Szegedy et al, “Inception-V4, Inception-ResNet and the Impact of Residual Connections on Learning”, arXiv 2016
MobileNet: Howard et al, “Efficient Convolutional Neural Networks for Mobile Vision Applications”, arXiv 2017

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 85

 Object Detection: Lots of variables ...
Backbone “Meta-Architecture” Takeaways
Network Two-stage: Faster R-CNN Faster R-CNN is slower
VGG16 Single-stage: YOLO / SSD but more accurate
ResNet-101 Hybrid: R-FCN
Inception V2 SSD is much faster but
Inception V3 Image Size not as accurate
Inception # Region Proposals
ResNet … Bigger / Deeper
MobileNet backbones work better
Huang et al, “Speed/accuracy trade-offs for modern convolutional object detectors”, CVPR 2017
Zou et al, “Object Detection in 20 Years: A Survey”, arXiv 2019
R-FCN: Dai et al, “R-FCN: Object Detection via Region-based Fully Convolutional Networks”, NIPS 2016
Inception-V2: Ioffe and Szegedy, “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift”, ICML 2015
Inception V3: Szegedy et al, “Rethinking the Inception Architecture for Computer Vision”, arXiv 2016
Inception ResNet: Szegedy et al, “Inception-V4, Inception-ResNet and the Impact of Residual Connections on Learning”, arXiv 2016
MobileNet: Howard et al, “Efficient Convolutional Neural Networks for Mobile Vision Applications”, arXiv 2017

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 86

 Instance Segmentation
Semantic Object Instance
Classification
Segmentation Detection Segmentation

CAT GRASS, CAT, DOG, DOG, CAT DOG, DOG, CAT

TREE, SKY

No spatial extent No objects, just pixels Multiple Object

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 87

 Object Detection:
Faster R-CNN

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 88

 Instance Segmentation: Mask Prediction

Mask R-CNN

Add a small mask

network that operates
on each RoI and
predicts a 28x28
binary mask

He et al, “Mask R-CNN”, ICCV 2017

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 89

 Mask R-CNN
Classification Scores: C
Box coordinates (per class): 4 * C

CNN Conv Conv

+RPN RoI Align

256 x 14 x 14 256 x 14 x 14 Predict a mask for

each of C classes

C x 28 x 28
He et al, “Mask R-CNN”, arXiv 2017

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 90

 Mask R-CNN: Example Mask Training Targets

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 91

 Mask R-CNN: Very Good Results!

He et al, “Mask R-CNN”, ICCV 2017

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 92

 Mask R-CNN
Also does pose

He et al, “Mask R-CNN”, ICCV 2017

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 93

Segment Parts and Objects
Segment Parts and Objects
Application: Parsing Historical Manuscripts
Application: Parsing Historical Manuscripts
https://ihdia.iiit.ac.in/Palmira/
Segmenting page instances
im2pages
 Open Source Frameworks
Lots of good implementations on GitHub!

TensorFlow Detection API:

https://github.com/tensorflow/models/tree/master/research/object_detection
Faster RCNN, SSD, RFCN, Mask R-CNN, ...

Detectron2 (PyTorch)
https://github.com/facebookresearch/detectron2
Mask R-CNN, RetinaNet, Faster R-CNN, RPN, Fast R-CNN, R-FCN, ...

Finetune on your own dataset with pre-trained models

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 99

 Beyond 2D Object Detection...

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 100

 3D Shape Prediction: Mesh R-CNN

Gkioxari et al., Mesh RCNN, ICCV 2019

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 101

 Object Detection + Captioning
= Dense Captioning

Johnson, Karpathy, and Fei-Fei, “DenseCap: Fully Convolutional Localization Networks for Dense Captioning”, CVPR 2016
Figure copyright IEEE, 2016. Reproduced for educational purposes.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 102

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 103
 Dense Video Captioning

Ranjay Krishna et al., “Dense-Captioning Events in Videos”, ICCV 2017

Figure copyright IEEE, 2017. Reproduced with permission.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 104

 Objects + Relationships = Scene Graphs

Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz,
Stephanie Chen et al. "Visual genome: Connecting language and vision using
crowdsourced dense image annotations." International Journal of Computer Vision 123,
no. 1 (2017): 32-73.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 105

 Scene Graph Prediction

Xu, Zhu, Choy, and Fei-Fei, “Scene Graph Generation by Iterative Message Passing”, CVPR 2017
Figure copyright IEEE, 2018. Reproduced for educational purposes.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 106

 3D Object Detection
2D Object Detection:
2D bounding box
(x, y, w, h)

3D Object Detection:
3D oriented bounding box
(x, y, z, w, h, l, r, p, y)

Simplified bbox: no roll & pitch

Much harder problem than 2D

object detection!
This image is CC0 public domain

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 107

 3D Object Detection: Simple Camera Model
A point on the image plane
corresponds to a ray in the 3D
3D ray space

A 2D bounding box on an image

is a frustrum in the 3D space
2D point
Localize an object in 3D:
The object can be anywhere in
camera
viewing frustrum the camera viewing frustrum!
image plane
camera

Image source: https://www.pcmag.com/encyclopedia_images/_FRUSTUM.GIF

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 108

 3D Object Detection: Monocular Camera

Faster R-CNN

- Same idea as Faster RCNN, but proposals are in 3D

- 3D bounding box proposal, regress 3D box parameters + class score
Chen, Xiaozhi, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel
Urtasun. "Monocular 3d object detection for autonomous driving." CVPR 2016.

Fei-Fei Li, Jiajun Wu, Ruohan Gao Lecture 9 - 109