Video has become ubiquitous on the Internet, TV, as well as personal devices. Recognition of video content has been a fundamental challenge in computer vision for decades, where previous research predominantly focused on recognizing videos using a predefined yet limited vocabulary. Thanks to the recent development of deep learning techniques, researchers in multiple communities are now striving to bridge videos with natural language in order to move beyond classification to interpretation, which should be regarded as the ultimate goal of video understanding. We will present recent advances in exploring the synergy of video understanding and language processing techniques.
Swan(sea) Song – personal research during my six years at Swansea ... and bey...
Video + Language
1. 2019/5/10
1
Video + Language
Jiebo Luo
Department of Computer Science
What Is Computer Vision?
• A major branch of Artificial Intelligence
• Vision is the process of discovering from images what
is present in the world, and where it is.
-- David Marr, Vision (1982)
Computer vision: 50 years of progress
6
feature engineering + model learning
𝐟 = 𝑓 𝐼 𝒚 𝑔 𝐟, 𝜃
deep le
𝒚 𝑔 𝐼, 𝜃 , wher
MSR-VTT
CCV
UCF101
HMDB51
ActivityNet
FCVID
Hollywood
Sports-1M
YouTube-
8M
1,000
10,000
100,000
1,000,000
10,000,000
10 100 1,000 10,000
#Example
#Class
Computer vision: 50 years of progress
7
COCO
KBK-1M
Open
Images
Visual
Genome
Caltech 101
Caltech 256
SUN
ImageNet
ImageNet…
Pascal
1,000
10,000
100,000
1,000,000
10,000,000
100,000,000
1 10 100 1,000 10,000 100,000
#Example
#Class
Flickr 8K
Flickr 30K
MPII-MD
M-VAD
MSVD
SBU
TGIF
YFCC 100M
Introduction
• Video has become ubiquitous on the Internet, TV, as well as
personal devices.
• Recognition of video content has been a fundamental challenge
in computer vision for decades, where previous research
predominantly focused on understanding videos using a
predefined yet limited vocabulary.
• Thanks to the recent development of deep learning techniques,
researchers in both computer vision and multimedia
communities are now striving to bridge video with
natural language, which can be regarded as the ultimate goal of
video understanding.
• We present recent advances in exploring the synergy of video
understanding and language processing, including video-
language alignment, video captioning, and video emotion
analysis.
From Classification to Description
Recognizing Realistic Actions from Videos "in the Wild"
UCF-11 to UCF-101
(CVPR 2009)
Similarity btw Videos Cross-Domain Learning
Visual Event Recognition in Videos by
Learning from Web Data
(CVPR2010 Best Student Paper)
Heterogeneous Feature Machine For Visual Recognition
(ICCV 2009)
1 5
6 7
9 10
2. 2019/5/10
2
From Classification to Description
11
Semantic Video Entity Linking
(ICCV2015)
Aligning Language Descriptions
with Videos
Iftekhar Naim, Young Chol Song, Qiguang Liu
Jiebo Luo, Dan Gildea, Henry Kautz
link
OverviewOverview
Unsupervised alignment of video with text
Motivations
Generate labels from data (reduce burden of manual labeling)
Learn new actions from only parallel video+text
Extend noun/object matching to verbs and actions
Unsupervised alignment of video with text
Motivations
Generate labels from data (reduce burden of manual labeling)
Learn new actions from only parallel video+text
Extend noun/object matching to verbs and actions
Matching Verbs to Actions
The person takes out a knife
and cutting board
Matching Nouns to Objects
[Naim et al., 2015]
An overview of the text and
video alignment framework
Hyperfeatures for ActionsHyperfeatures for Actions
High-level features required for alignment with text
→ Motion features are generally low-level
Hyperfeatures, originally used for image recognition extended
for use with motion features
→ Use temporal domain instead of spatial domain for vector
quantization (clustering)
High-level features required for alignment with text
→ Motion features are generally low-level
Hyperfeatures, originally used for image recognition extended
for use with motion features
→ Use temporal domain instead of spatial domain for vector
quantization (clustering)
Originally described in “Hyperfeatures:
Multilevel Local Coding for Visual Recog‐
nition” Agarwal, A. (ECCV 06), for images Hyperfeatures for actions
Hyperfeatures for ActionsHyperfeatures for Actions
From low-level motion features, create high-level
representations that can easily align with verbs in text
From low-level motion features, create high-level
representations that can easily align with verbs in text
Cluster 3
at time t
Accumulate over
frame at time t
& cluster
Conduct vector
quantization
of the histogram
at time t
Cluster 3, 5, …,5,20
= Hyperfeature 6
Each color code
is a vector
quantized
STIP point
Vector quantized
STIP point histogram at time t
Accumulate clusters over
window (t‐w/2, t+w/2]
and conduct vector
quantization
→ first‐level hyperfeatures
Align hyperfeatures
with verbs from text
(using LCRF)
Latent-variable CRF AlignmentLatent-variable CRF Alignment
CRF where the latent variable is the alignment
N pairs of video/text observations {(xi, yi)}i=1 (indexed by i)
Xi,m represents nouns and verbs extracted from the mth sentence
Yi,n represents blobs and actions in interval n in the video
Conditional likelihood
conditional probability of
Learning weights w
Stochastic gradient descent
CRF where the latent variable is the alignment
N pairs of video/text observations {(xi, yi)}i=1 (indexed by i)
Xi,m represents nouns and verbs extracted from the mth sentence
Yi,n represents blobs and actions in interval n in the video
Conditional likelihood
conditional probability of
Learning weights w
Stochastic gradient descent
where feature function
More details in Naim et al. 2015 NAACL Paper ‐
Discriminative unsupervised alignment of natural language instructions with corresponding video segments
N
11 14
16 17
18 19
3. 2019/5/10
3
Experiments: Wetlab DatasetExperiments: Wetlab Dataset
RGB-Depth video with lab protocols in text
Compare addition of hyperfeatures generated from motion
features to previous results (Naim et al. 2015)
Small improvement over previous results
Activities already highly correlated with object-use
RGB-Depth video with lab protocols in text
Compare addition of hyperfeatures generated from motion
features to previous results (Naim et al. 2015)
Small improvement over previous results
Activities already highly correlated with object-use
Detection of objects in 3D space
using color and point‐cloud
Previous results
using object/noun
alignment only
Addition of different types
of motion features
2DTraj: Dense trajectories
*Using hyperfeature window size w=150
Experiments: TACoS DatasetExperiments: TACoS Dataset
RGB video with crowd-sourced text descriptions
Activities such as “making a salad,” “baking a cake”
No object recognition, alignment using actions only
Uniform: Assume each sentence takes the same amount of time over the entire sequence
Segmented LCRF: Assume the segmentation of actions is known, infer only the action labels
Unsupervised LCRF: Both segmentation and alignment are unknown
Effect of window size and number of clusters
Consistent with average
action length: 150 frames
RGB video with crowd-sourced text descriptions
Activities such as “making a salad,” “baking a cake”
No object recognition, alignment using actions only
Uniform: Assume each sentence takes the same amount of time over the entire sequence
Segmented LCRF: Assume the segmentation of actions is known, infer only the action labels
Unsupervised LCRF: Both segmentation and alignment are unknown
Effect of window size and number of clusters
Consistent with average
action length: 150 frames
*Using hyperfeature
window size w=150
*d(2)=64
Experiments: TACoS DatasetExperiments: TACoS Dataset
Segmentation from a sequence in the dataset Segmentation from a sequence in the dataset
Crowd‐sourced descriptions
Example of text and video alignment generated
by the system on the TACoS corpus for sequence s13‐d28
Image Captioning with Semantic
Attention (CVPR 2016)
Quanzeng You, Jiebo Luo
Hailin Jin, Zhaowen Wang and Chen Fang
Image Captioning
• Motivations
– Real-world Usability
• Help visually impaired people, learning-impaired
– Improving Image Understanding
• Classification, Objection detection
– Image Retrieval
1. a young girl inhales with the intent of blowing
out a candle
2. girl blowing out the candle on an ice cream
1. A shot from behind home
plate of children playing
baseball
2. A group of children playing
baseball in the rain
3. Group of baseball players
playing on a wet field
Introduction of Image Captioning
• Machine learning as an approach to solve the problem
20 21
22 23
24 25
4. 2019/5/10
4
Overview
• Brief overview of current approaches
• Our main motivation
• The proposed semantic attention model
• Evaluation results
Brief Introduction of Recurrent Neural Network
• Different from CNN
11),( ttttt BhAxhxfh
tt Chy
• Unfolding over time
Feedforward network
Backpropagation through time
Inputs
Hidden Units
Outputs
xt ht-1
ht
yt
Inputs
Hidden Units
Outputs
C
yt
Inputs
Hidden Units
Inputs
Hidden Units
B
B
A
A
A B
t-1
t-2
Applications of Recurrent Neural Networks
• Machine Translation
• Reads input sentence “ABC” and produces “WXYZ”
Decoder RNNEncoder RNN
Encoder-Decoder Framework for Captioning
• Inspired by neural network based machine
translation
• Loss function
N
t
tt wwIwp
IwpL
1
10 ),,,|(log
)|(log
Our Motivation
• Additional textual information
– Own noisy title, tags or captions (Web)
Our Motivation
• Additional textual information
– Own noisy title, tags or captions (Web)
– Visually similar nearest neighbor images
26 27
28 29
31 32
5. 2019/5/10
5
Our Motivation
• Additional textual information
– Own noisy title, tags or captions (Web)
– Visually similar nearest neighbor images
– Success of low-level tasks
• Visual attributes detection
Image Captioning with Semantic Attention
• Big idea
First Idea
• Provide additional knowledge at each input node
• Concatenate the input word and the extra attributes K
• Each image has a fixed keyword list
)],,([),( 11 tktttt hbKWwfhxfh
Visual Features: 1024
GoogleNet
LSTM Hidden states: 512
Training details:
1. 256 image/sentence
pairs
2. RMS-Prob
Using Attributes along with Visual Features
• Provide additional knowledge at each input node
• Concatenate the visual embedding and keywords for h0
];[),( 10 bKWvWhvfh kiv
Attention Model on Attributes
• Instead of using the same set of attributes at every
step
• At each step, select the attributes (attention)
m mtmt kKwatt ),(
)softmax VK(wT
tt
))],,(;([),( 11 tttttt hKwattxfhxfh
Overall Framework
• Training with a bilinear/bilateral attention model
ht
pt
xt
v
{Ai}
Yt~
𝜑
𝜙
RNN
Image
CNN
AttrDet 1
AttrDet 2
AttrDet 3
AttrDet N
t = 0
Word
33 34
35 36
38 39
6. 2019/5/10
6
Visual Attributes
• A secondary contribution
• We try different approaches
Performance
• Examples showing the impact of visual attributes on captions
Performance on the Testing Dataset
• Publicly available split
Performance
• MS-COCO Image Captioning Challenge
Captioning with Emotion and Style A Simple Framework
40 41
42 43
45 46
7. 2019/5/10
7
Examples Examples
Integrating Scene Text and Visual Appearance
for Fine-Grained Image Classification Xiang Bai et al.
49
TGIF: A New Dataset and Benchmark on
Animated GIF Description
Yuncheng Li, Yale Song, Liangliang Cao, Joel Tetreault,
Larry Goldberg, Jiebo Luo
Overview Comparison with Existing Datasets
47 48
49 50
51 52
8. 2019/5/10
8
Examples
a skate boarder is
doing trick on his skate
board.
a gloved hand opens to
reveal a golden ring.
a sport car is swinging on
the race playground
the vehicle is moving fast
into the tunnel
Contributions
• A large scale animated GIF description dataset for promoting image
sequence modeling and research
• Performing automatic validation to collect natural language descriptions
from crowd workers
• Establishing baseline image captioning methods for future
benchmarking
• Comparison with existing datasets, highlighting the benefits with
animated GIFs
In Comparison with Existing Datasets
• The language in our dataset is closer to common language
• Our dataset has an emphasis on the verbs
• Animated GIFs are more coherent and self contained
• Our dataset can be used to solve more difficult movie description
problem
Machine Generated Sentence Examples
Machine Generated Sentence Examples Machine Generated Sentence Examples
53 54
55 56
57 58
9. 2019/5/10
9
Comparing Professionals and Crowd-workers
Crowd worker: two people are kissing on a boat.
Professional: someone glances at a kissing couple
then steps to a railing overlooking the ocean an
older man and woman stand beside him.
Crowd worker: two men got into their car
and not able to go anywhere because the
wheels were locked.
Professional: someone slides over the
camaros hood then gets in with his partner
he starts the engine the revving vintage
car starts to backup then lurches to a halt.
Crowd worker: a man in a shirt and tie sits beside a person who is
covered in a sheet.
Professional: he makes eye contact with the woman for only a second.
More: http://beta-
web2.cloudapp.net/ls
mdc_sentence_comp
arison.html
Movie Descriptions versus TGIF
• Crowd workers are encouraged to describe the major visual content
directly, and not to use overly descriptive language
• Because our animated GIFs are presented to crowd workers without
any context, the sentences in our dataset are more self-contained
• Animated GIFs are perfectly segmented since they are carefully
curated by online users to create a coherent visual story
Code & Dataset
• Yahoo! webscope (Coming soon!)
• Animated GIFs and sentences
• Code and models for LSTM baseline
• Pipeline for syntactic and semantic validation to collect natural
languages from crowd workers
Image Captioning with Unpaired Data
• Supervised and semi-supervised captioning methods rely on paired
training data, which are expensive to acquire (i.e., the bottleneck)
• Image captioning trained with unpaired data is a promising direction
because unpaired training data are much easier to obtain
Supervised captioning
images captions
Semi-supervised captioning
images captions
Unpaired captioning
images captions
(a) (b) (c)
Image Captioning with Unpaired Data
• How to align images with sentences?
… …
Little girl looking
down at leaves
with her bicycle
with training
wheels parked
next to her.
Image Captioning with Unpaired Data
• How to align images with sentences?
… …
Little girl looking
down at leaves
with her bicycle
with training
wheels parked
next to her.
Little girl looking
down at leaves
with her bicycle
with training
wheels parked
next to her.
Object
Detection
59 60
61 62
63 64
10. 2019/5/10
10
Image Captioning with Unpaired Data
Bench, bike, girl.
How to
generate a
sentence?
Image Captioning with Unpaired Data
Bench, bike, girl.
…
A young child in a park next to a red
bench and red bicycle that as training
wheels.
A woman that is standing in front of an oven.
Real sentences
Discriminator Real/Fake
Image Captioning with Unpaired Data Image Captioning with Unpaired Data
Unpaired Image Captioning
detected concepts person, tree, helmet
con2sen a person in a helmet is riding a tree .
feat2sen
a person wearing a helmet is skateboarding on a ramp
.
adv a man riding a skateboard down a street .
adv + con a person on a court with a racket
adv + con + im a man riding a skateboard on a ramp .
Ours without init a man riding a surfboard on top of a wave .
Ours a man on a skateboard performing a trick .
BRNN a man riding a skateboard down the side of a ramp
detected concepts boat, watercraft
con2sen a group of people in a boat with a lighthouse .
feat2sen a boat that is floating in the water .
adv a man riding a wave on top of a surfboard .
adv + con a large boat that is in the water .
adv + con + im a large boat in the middle of a lake .
Ours without init a boat that is sitting in the water
Ours a view of a boat in the ocean
BRNN a boat floating on top of a body of water
Unpaired Image Captioning
detected concepts boat
con2sen a boat that is floating in the water .
feat2sen a boat that is floating in the water .
adv a small boat is floating in the water .
adv + con a boat that is floating in the water .
adv + con + im a boat that is parked in the water .
Ours without init a boat that is sitting in the water .
Ours a boat is docked in a harbor .
BRNN a boat floating on top of a body of water
detected concepts animal, frame
con2sen a stuffed animal sitting on top of a frame .
feat2sen
a group of people standing around a table with a
picture of a large piece of cloth and
adv a man riding a wave on top of a surfboard .
adv + con a group of people that are standing on the beach
adv + con + im a person riding a surf board on the waves
Ours without init a child is flying a kite on the beach .
Ours a group of people surfing on a wave .
BRNN a man standing on top of a sandy beach
65 66
67 68
69 70
11. 2019/5/10
11
Sports Video Captioning
by Attentive Motion Representation
based Hierarchical Recurrent Neural Networks
Mengshi Qi1, Yunhong Wang1, Annan Li1, Jiebo Luo2
1Beijing Advanced Innovation Center for Big Data and Brain Computing,
School of Computer Science and Engineering, Beihang University
2Department of Computer Science, University of Rochester
Introduction
Sports video captioning is a task of automatically generating a
textual description for sports events (e.g., soccer, basketball or
volleyball games).
Soccer Basketball Volleyball
Problem Definition Challenges
How to design an effective sports video captioning system?
conventional video captioning methods can only generate coarse descriptions,
and overlook the motion details of players’ actions and group-level activities
How to capture fine-grained motion information of sports
players?
player’s articulated movements/pose estimation/motion trajectory
the key player’s action is critical in a sports event
How to build a dataset that focuses on sports video captioning?
different sports games
Contributions
We introduce a novel framework for sports video captioning with
attentive motion representation-based hierarchical recurrent neural
networks.
A motion representation module is employed to extract player’s pose
and trajectory information, where we capture semantic attributes from
player’s skeletons and cluster trajectories from dynamic movements.
We annotate a new dataset called Sports Video Captioning Dataset-
Volleyball .
Framework
71 72
73 74
75 76
12. 2019/5/10
12
(1) Action Proposal Module
Given a video, retrieving and localizing temporal segments that
likely contain spatio-temporal group events (i.e. attack, defend)
or individual actions (i.e. spiking, passing) is the first task.
Deep Action Proposals (DAP) method [1]
Motion Representation Module
(1) Pose Attribute Detection Part
(2) Trajectory Clustering Part
Pose Attribute Detection Part
Extract pose-based CNN (Convolutional Neural Network) feature [2] from each
body part of a player
Build an attribute vocabulary from the annotated sentences in the dataset (e.g.,
UCF-101 [3] , Volleyball [4]), and use the top 𝑘 high-frequency words
Trajectory Clustering Part
Extracting the dense point trajectories 𝑇𝑟𝑎 𝑇𝑟𝑎 , … , 𝑇𝑟𝑎 for a sequence of
frames by trajectory-pooled deep-convolutional descriptors [5] , where 𝑀 is the
number of trajectories
Then we partition all the detected trajectories into groups by computing the
affinity matrix between each trajectory pair and utilizing the graph clustering
method [6], we can obtain 𝑚 clusters.
Databases
General Video Captioning:
1) Microsoft Video Description Dataset (MSVD)
2) MSR Video-to-Text Dataset (MSR-VTT)
Sports Video Captioning:
1) Sports Video Captioning Dataset-Volleyball (SVCDV)
Sports Video Captioning Dataset-Volleyball
(SVCDV)
55 videos with 4,830 short clips from Olympic Volleyball Games
About 44,436 sentences, 9.2 sentences on average of each
clip
Verb ratio is nearly 16.2%, 1.72 verbs per sentences
9 individual action labels:
waiting/setting/digging/falling/spiking/blocking/jumping/moving/standing
8 group activity labels:
right set/right spike/right pass/right winpoint/left winpoint/left pass/left
spike/left set
77 78
79 81
86 87
13. 2019/5/10
13
General Video Captioning Task
Table 1. Captioning Performance on MSV and
MSR-VTT Datasets
General Video Captioning Task
Sports Video Captioning and Components
Analysis
Table 2. Captioning Performance on SVCDV
Sports Video Captioning
Demo Video Video Re-localization (ECCV 2018)
• How to localize the part semantically relevant to the clip on the
left in the long video on the right?
88 89
90 91
92 95
14. 2019/5/10
14
Video Re-localization
Possible solutions
• Copy detection
–The contents can be largely different in both foreground and
background
• Action localization
–Only one sample is available
Video Re-localization
• We propose a matching framework
Cross Gating & Bilinear Matching
• We propose
• Cross gating to remove irrelevant contents
• Bilinear matching to capture the interactions between the query and
reference
Video Re-localization
• Results
Shuffleboard
query length = 14
reference length = 49
Ground truth 17 27
Ours 17 27
Frame 15 23
Video 17 19
Pole vault
query length = 15
reference length = 49
Ground truth 10 37
Ours 10 38
Frame 20 32
Video 17 38
query
reference
query
reference
Video Re-localization Localizing Language in Video (AAAI 2019)
• How to localize the part semantically relevant to the clip on the
left in the long video on the right?
96 97
98 99
100 101
15. 2019/5/10
15
Localizing Language in Video (AAAI 2019) Localizing Language in Video (AAAI 2019)
Localizing Language in Video (AAAI 2019) Localizing Language in Video (AAAI 2019)
Localizing Language in Video (AAAI 2019) Increasing Attention (CVPR 2019)
102 103
104 105
106 131