video + language: where does domain knowledge fit in?
TRANSCRIPT
Video + Language: Where Does Domain Knowledge Fit in?Jiebo LuoDepartment of Computer Science
July 10, 2016
Keynote@2016 IJCAI Workshop on Semantic Machine Learning
Domain Knowledge in Machine Learning
• Domain knowledge is used frequently in ML applications (sometimes without knowing that you are actually doing it)– A good example is feature extraction. What features to use?– Other uses include objective function, parameter selection– Even in deep learning (architecture, learning rate, etc.) – Certainly probabilistic graphical models (including priors)– Data cleaning (yes!)
• Context models encode domain knowledge– Spatial context (e.g., in computer vision)– Temporal context (e.g., in sequence analysis)– Social context (e.g., in social media data mining)
• We will focus on some less obvious, more sophisticated forms of domain knowledge, especially in the area of “vision and language”, an emerging fertile ground in machine learning
IEEE Signal Processing, 2005
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.
Video Growth
• CISCO Trends
6
SEMANTICS <> USER INTENT
SEMANTICS IN LARGE SCALE VIDEO SEARCH & RETRIEVAL
SEMANTIC VIDEO ENTITY LINKING
Users
User Modeling
User Content Understanding
Learning From User
Tags
Yuncheng Li, Xitong Yang, Jiebo Luo
Semantic Video Entity Linking
Semantic Video Entity Linking
• Motivations to use visual content1. Video entity linking is very challenging with only
title & descriptions, especially for UGC2. Video entity linking must be of high quality3. The visual content of a video truly represents the
user intent for video watching and sharing
Semantic Video Entity Linking
Semantic Video Entity Linking
• Sequence-to-Set matchingKey Frame Sequence
…
Representative Image Set
…
Match?
Semantic Video Entity Linking
• Train Data: Triplets (Keyframes, Images, Label)Key Frame Sequence
…
Representative Image Set
Yes
Key Frame Sequence
No
Representative Image Set
Semantic Video Entity Linking
• Goal: Keyframe-Image distance (metric)Key Frame Sequence
…
Representative Image Set
…
Semantic Video Entity Linking
• Challenge: Not all pairs are true matchesKey Frame Sequence
…
Representative Image Set
…
Yes
Semantic Video Entity Linking
• Solution: Multiple Instance Metric Learning– EM like algorithm:
• E-Step: infer the true matches• M-Step: Retrain the metric
• More Challenges– Overwhelming noise (variability & ambiguity) in
both visual content• Solution
– Structured constraints to suppress noise
Semantic Video Entity Linking
Title: “Jason Taylor Career Highlights”
Semantic Video Entity Linking
• Experiments– Dataset: 1920 videos– Labels: Amazon Mechanical Turk
Semantic Video Entity Linking
We propose two constraints to guide the video entity linking process:1. Temporal Smoothness: the entity occurrence (matches with any of the representative images) is smooth over time 2. Representativeness Smoothness: in order to reduce the significant irrelevant information
Conclusions
• Metric Learning helps by adapting to different topics and domains
• Structured constraints are important for suppressing noises
• Future work includes integrating the video metadata information and building entity integrated applications, e.g., video spam detection
Semantic Video Entity Linking. An Application: Video Spam Removal
• “guardians of the galaxy full movie”– Let’s watch the movie
Unsupervised Alignment of Actions in Video with Text Descriptions
Y. Song, I. Naim, A. Mamun, K. Kulkarni, P. SinglaJ. Luo, D. Gildea, H. Kautz
Overview• 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 ActionsThe person takes out a knife
and cutting board
Matching Nouns to Objects
[Naim et al., 2015]
An overview of the text andvideo alignment framework
Hyperfeatures 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)
Originally described in “Hyperfeatures:Multilevel Local Coding for Visual Recog-nition” Agarwal, A. (ECCV 06), for images Hyperfeatures for actions
Hyperfeatures for Actions• From low-level motion features, create high-level
representations that can easily align with verbs in text
Cluster 3at time t
Accumulate overframe at time t
& cluster
Conduct vectorquantizationof the histogramat time t
Cluster 3, 5, …,5,20= Hyperfeature 6
Each color codeis a vectorquantizedSTIP point
Vector quantizedSTIP point histogram at time t
Accumulate clusters overwindow (t-w/2, t+w/2]and conduct vectorquantization→ first-level hyperfeatures
Align hyperfeatureswith verbs from text
(using LCRF)
Latent-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
wherefeature function
More details in Naim et al. 2015 NAACL Paper -Discriminative unsupervised alignment of natural language instructions with corresponding video segments
N
Experiments: 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
Detection of objects in 3D spaceusing color and point-cloud
Previous resultsusing object/nounalignment only
Addition of different typesof motion features
2DTraj: Dense trajectories*Using hyperfeature window size w=150
Experiments: 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
*Using hyperfeaturewindow size w=150
*d(2)=64
Experiments: TACoS Dataset
• Segmentation from a sequence in the dataset
Crowd-sourced descriptionsExample 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 LuoHailin 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 candle2. 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
Model sentence
1. A young girl inhales with the intent of blowing out a candle.2. A young girl is preparing to blow out her candle.3. A kid is to blow out the single candle in the bowl of birthday goodness.4. Girl blowing out the candle on an ice-cream5. A little girl is getting ready to blow out a candle on a small dessert.
1. A shot from behind home plate of children playing baseball2. A group of children playing baseball in the rain3. Group of baseball players playing on a wet field4. A batter leaning back so they don’t get hit by a ball5. A group of young boys playing baseball in the rain
1. A girl in a park area flies a multi-colored kite.2. A girl flying a kit in the sky3. A young woman flying a rainbow colored kite.4. A person in a large field flying a kite in the sky.5. A woman looks up at her colorful sailing kite.
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
...
Convolutional Neural NetworkInputs
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
ttt wwIwp
IwpL
110 ),,,|(log
)|(log
...
Convolutional Neural Network
w1
wstart
w2
w1
wN
wN-1
wend
wN
Image
Recurrent Neural Network
...
Convolutional Neural Network
#Start#
Some
riverSome
elephants .
Some elephants roaming around on a river bank.
bank
bank
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
• Main idea
RNN ...
CNN
attention
waveridingman
surfboardoceanwatersurfersurfingpersonboard
0
0.1
0.2
0.3
surfboardwavesurfing
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 GoogleNetLSTM Hidden states: 512
Training details:1. 256 image/sentence pairs 2. RMS-Prob
...
Convolutional Neural Network
w1
wstart
w2
w1
wN
wN-1
wend
wN
Retrieve Tags, titles, descriptions
from weak annotated images
Feature extraction
K
Keywords, key-phrase
Image
Recurrent Neural Network
Using Attributes along with Visual Features
• Provide additional knowledge at each input node
• Concatenate the visual embedding and keywords for h0
];[),( 10 bKWvWhvfh kiv +== −
...
Convolutional Neural Network
w1
wstart
w2
w1
wN
wN-1
wend
wN
Retrieve Tags, titles, descriptions
from weak annotated images
Feature extraction
K
Keywords, key-phrase
Image
h0
Recurrent Neural Network
Attention Model on Attributes
• Instead of using the same set of attributes at every step• At each step, select the attributes
∑= m mtmt kKwatt α),(
)softmax VK(wTtt =α
))],,(;([),( 11 −− == tttttt hKwattxfhxfh
...
CNN
w1
wstart
w2
w1
wN
wN-1
wend
wN
Retrieve Tags, titles, descriptions
from weak annotated images
Feature extraction
K
Keywords, key-phrase
RNN
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
RNN ...
CNN
attention
waveridingman
surfboardoceanwatersurfersurfingpersonboard
0
0.1
0.2
0.3
surfboardwavesurfing
Visual Attributes
• A secondary contribution• We try different approaches
vase flowers bathroom table glass sink blue small white clear
k-NN
sitting table small many little glass different flowers vase shown
Multi-label Ranking
vase flowers table glass sitting kitchen water room white filled
FCN
Performance
• Examples showing the impact of visual attributes on captions
Google NIC
Top-5 visual attributes
ATT-FCN
a white plate topped with a variety of food.
a plate with a sandwich and french fries.
plate broccoli fries food french
a baby with a toothbrush in its mouth.
a baby is eating a piece of paper.
teeth brushing toothbrush holding baby
a traffic light is on a city street.
a street with cars and a clock tower.
street sign cars clock traffic
a yellow and black train on a track.
a train traveling down tracks next to a building.
train tracks clock tower down
a close up of a plate of food on a table.
a table topped with a cake with candles on it.
a teddy bear sitting on top of a chair .
a white teddy bear sitting next to a stuffed animal .
a person is holding colorful umbrella.
a black umbrella sitting on top of a sandy beach .
a woman is holding a cell phone in her hand .
a woman holding a pair of scissors in her hands .
cake table plate sitting birthday
teddy cat bear stuffed white
umbrella beach water sitting boat
woman bathroom her scissors man
Performance on the Testing Dataset
• Publicly available split
Performance• MS-COCO Image Captioning Challenge
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
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
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/lsmdc_sentence_comparison.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
Where is CV (AI) in 2016?Winter 2002 Fall 2003 Summer 2008
Image/Video captioning看图识字 看图说话
ThanksQ & A
Google***
BaiduSogou
BingXiaoIce
Are you smarter than a 5th grader?
What doe it take to go from a 5-year old to a 5th grader?1. Learning from “small data”2. Unsupervised learning3. Transfer learning4. Integration of domain
knowledge or experience
Visual Intelligence & Social Multimedia Analyticswww.cs.rochester.edu/u/jluo
Questions?