Daily

In cooperation with

Computer Vision NewsThe Magazine of The Algorithm community

A publication by

Today’s Picks

HighlightsWomen in Computer VisionPresentations

THURSDAYJune 30

CVPR Expo

Morning:09:00AM to 10:18 AMNon-Rigid Reconstruction and Motion AnalysisOral session O4-1B-17:Full Flow: Optical Flow Estimation By Global Optimization Over Regular GridsPage 40 of the Pocket Guide and presentation 17 on CVPR’s website

Afternoon:1:45PM to 2:50PMLearning and CNN ArchitecturesOral session O4-2A-2Structural-RNN: Deep Learning on Spatio-Temporal GraphsPage 43 of the Pocket Guide and presentation 2 on CVPR’s website

3D Shape ReconstructionOral session O4-2B-17SVBRDF-Invariant Shape and Reflectance Estimation From Light-Field CamerasPage 43 of the Pocket Guide and presentation 17 on CVPR’s website

3:45PM to 2:50PM3D, Stereo, Matching, and Saliency EstimationSpotlight session S4-3B-45FANNG: Fast Approximate Nearest Neighbour GraphsPage 44 of the Pocket Guide and presentation 45 on CVPR’s website

People and FacesSpotlight session S4-3A-34Fits Like a Glove: Rapid and Reliable Hand Shape PersonalizationPage 44 of the Pocket Guide and presentation 34 on CVPR’s website

2CVPR Daily: Thursday

For Today, Thursday 30

David’s Picks

David Fleet is Professor of ComputerScience at the University of Toronto, andChair of Computer and MathematicalSciences at the University of TorontoScarborough. David accepted to share hispicks for today with CVPR Daily. Here areDavid’s picks for Thursday 30. Don’t missthem!

“I have not read these papers in any detail yet, but I am

interested in learning more”

Welcome 3CVPR Daily: Thursday

Dear Reader,

This is the last issue of CVPR Daily for this year. Thus, even though CVPR 2016is still in full swing, we would like to think for a moment about what we'veexperienced here in Las Vegas.

Everyone agrees about the exceptional level of the technical program,including the award-winning works:

1) Deep Residual Learning for Image Recognition, by Kaiming He, Xiangyu Zhang,Shaoqing Ren and Jian Sun;2) Sublabel-Accurate Relaxation of Nonconvex Energies, by Thomas Möllenhoff,Emanuel Laude, Michael Moeller, Jan Lellmann and Daniel Cremers;3) Structural-RNN: Deep Learning on Spatio-Temporal Graphs; by Ashesh Jain,Amir R. Zamir, SilvioSavarese and Ashutosh Saxena.

Of course there are many other excellent papers featured at the conference.Through hundreds of presentations, we have discovered the amazing talent ofmany outstanding researchers. In its 4 first-ever issues, CVPR Daily has doneits best to uncover, at least partially, all this talent and expose the work ofmore than 100 scientists.

One of the goals of this magazine is to put some name and stories with themany faces in the crowd. We deliberately chose to highlight the diversity ofpeople who are part of our research community, in particular introducing youto the many brilliant women at this year's CVPR.

Finally, one word about CVPR's growth. We had 2800 attendees in 2015; weare at 3609 attendees this year. That makes a year over year growth of nearly30%. This exceptional progression enables us to look forward to the future ofour community with full confidence and enthusiasm.

We hope the last days of CVPR 2016 are fruitful for you. We believe this weekhas created countless channels of communication and strengthened evenmore the bonds of the international computer vision community. Wewholeheartedly hope to see you all again at CVPR 2017!

Olga Russakovsky Ralph AnzarouthPress Chair, CVPR 2016 Editor, Computer Vision NewsRobotics Institute: Carnegie Mellon University Marketing Manager, RSIP Vision

CVPR Daily: Danna, what is yourcurrent position?

Danna Gurari: I am currently apost-doctoral fellow with KristenGrauman at University of Texasat Austin.

CVPR Daily: What is the subjectof your work?

Danna: I’ve been working at lotwith microscopy images as wellas everyday images. I look athow to actively involve humansto solve tasks such as objectsegmentation and visualquestion and answering.

CVPR Daily: Why did you decideto dedicate your career tocomputer vision?

Danna: I think that’s just aninstinctual love.

CVPR Daily: Do scientists followtheir instincts too?

Danna: I think it better besomething you love when itcomes to what you do every daywhen you wake up and what’sgoing to motivate you to do yourwork.

CVPR Daily: Do you love whatyou do?

Danna: I love what I do.

CVPR Daily: Is it difficult to be awoman in computer vision? Ifyes, why?

Danna: I would say it’s been a loteasier for me in academia than inindustry. I worked in industry forfive years, and I found that I wastypically alone as a female. Thatwas harder.

In academia, I’ve been veryfortunate because my advisor formy PhD as well as my currentpostdoc advisor are female. Ifind, in general, the communityreally fosters and creates a safeplace for people my age who arefemales.

4 Women in Computer Vision

Danna Gurari

CVPR Daily: Thursday

“I love what I do”

CVPR Daily: That’s very interestingbecause several times I’ve heardwomen saying that being the onlywoman was a limitation. But if Iwere the only man among manywomen, I wouldn’t feel isolated.Why is it so different?

Danna: I would say the lifechallenges are different for menand women. Long term, a lot of usin academia think of our academiccareer and how to balance thatwith a home life. I also think thereare other female aspects. There’ssomething intangible about havingsomeone you can relate to thatway. When I think about a rolemodel, typically I feel moreconnected to a female. That storymatches mine better.

As a counterpoint, my father was afaculty in computer science. Hewas one of my main mentors. He’sbeen a strong male presence in mylife, especially in computer science.

Women in Computer Vision 5CVPR Daily: Thursday

“I worked in industry for five years, and I

found that I was typically alone as a female. That was

harder”

CVPR Daily: What would you like toachieve in your career?

Danna: I would like to have my workhelp save lives. I work a lot in themicroscopy and the medical imagingspace. My greatest wish is thatsomething I do will in some way helpsave someone’s life.

CVPR Daily: What are the chances thatyou give yourself to be able to fulfil thishope one day?

Danna: I think it’s quite possible. As amentor, I can be a voice that creates amovement for this area. Whether it’sme or someone I teach, both feelequally good.

CVPR Daily: You may have noticed thatCVPR Daily highlights the diversity ofpeople who are part of our researchcommunity, in particular introducingthe many brilliant women at this year'sCVPR. What do you think of that?

Danna: I trust those that are makingthe decisions. It’s interesting to me,and I don’t think I have the full picture.I just trust that there’s some value intrying to put a lot of female faces outthere. Maybe it’s just a small thing ofmaking the younger females coming toCVPR feel like they fit in. It could makea long term impact on females goingforward. I’m not really sure of theintention, but I can imagine goodoutcomes from it.

6 Women in Computer VisionCVPR Daily: Thursday

David’s Picks for Thu30 2

Editorial 3

Women in Computer Vision: Danna Gurari 4Angels Rates Borras 12

Summary 6

CVPR Expo:

Qihoo 360 7TuSimple 10MorpX 11

Presentations:M. Blaha and A. Richard 8

Guy Rosman - CSAIL/MIT 14

Highlights 9

CVPR DailyEditor: Ralph AnzarouthPublisher: RSIP VisionMany thanks to Allison Dienstman

Copyright: CVPR & RSIP VisionAll rights reservedUnauthorized reproductionis strictly forbidden.

CVPR Daily’s editorial choices are fully independent from CVPR

Dear reader,

Would you like to keepreading great computervision stories? Subscribeto Computer Vision Newsand receive it for free inyour mailbox everymonth!

FREE

SUBSCRIPTION

Subscription Form(click here, it’s free)

“I would like to have my work help save lives”

CVPR Expo 7CVPR Daily: Thursday

Shuicheng and Cheong invited me to see something very special at theQihoo 360 booth and I was not disappointed. Follow the guide!

They showed me a very good facial analysis system. Very accurate, stable,and energy efficient. Accurate means that if you stand in front of thecamera, the position or the important points will be very accurate.

Second, it will be very stable. Stable also means when you speak, all thepoints will stick on your face with markers. It is also very robust, in thesense that when only a part of the face is visible, even only one quarter ofit, these points are still very accurate and stable.

With this kind of functions, many things become possible: make a facelook thinner or enlarge the eyes to make it look more beautiful. In Asianculture, Shuicheng told me, thin face and big eyes make one’s aspectbeautiful. You can also perform analysis, for example, of the age orgender, whether you wear glasses, how beautiful you are, etc. Yourhumble servant tried this out, see the results in the bottom right image!

Qihoo 360

Mask transfer effect

This is what Qihoo’s software says about me! I look 2 years younger than I am,

I wear glasses, I am a gentleman (of course,I am Italian!) and my face is smiling

The Qihoo/360 AI Institute claimsa face SDK which is very stable inregards to head movement,expression changes and posechanges, as well as robust topartial visibility of the face. Thesefeatures are critical for facemanipulations/editing like masktransfer (see image below).

The Large-Scale Semantic 3D Reconstruction project was presentedboth at the orals and as a poster. What it does is automatic generationof semantically annotated 3D city models, starting from large imagecollections. The idea of semantic 3D reconstruction is to recover thegeometry of an observed scene while at the same time alsointerpreting the scene in terms of semantic object classes (such asbuildings, vegetation etc.).

While jointly inferring 3D shapes and semantic classes deliversappealing results, up-to-date methods are memory-hungry andcomputationally expensive. This approach demonstrates how to scalesemantic 3D reconstruction up to be able to reconstruct largegeographic areas such as cities. The resulting superior 3D models allowfor advanced reasoning tasks such as urban planning, navigation,physics-based simulations and the like.

This method is not only limited to 3D city modelling: possible furtherapplications include a number of multi-class segmentation problems,e.g. in the field of medical imaging. This six-teamer work is available infull here.

CVPR Daily: Thursday

8 Presentations

Maros Blaha and Audrey Richard

CVPR Daily: Thursday

This is what it looks when many of the 3,609 CVPR attendees, coming from countlesscountries and from all the continents, have lunch together. CVPR Daily owes this lovelyimage of friendship and comradery to the kindness of Dave Rushton-Smith of Google.

Highlights 9CVPR Daily: Thursday

Is There a Scientist in the Room?

Spare5

Besides distributing T-shirts, at theSpare5 booth they talk about theirservice enabling machine learningteams to generate accurate and diverseannotations on their datasets to train,validate, and test their algorithms.Their Training Data as a Service platform is a complete solution foracquiring and annotating datasets and building ground truth at scale. Iwas told there that leaders in machine learning rely on Spare5'sautomated platform for efficiencies in accuracy, specificity, scale, cost,and reduced level of effort in computer vision and natural languageprocessing applications.

If you ask why the company is called TuSimple, this must havesomething to do with “Technology made simple”. What they do isADAS (Advanced Driver Assistance System), autonomous driving,and web-scale image recognition applications.

We asked them about that and Amanda Song explained that theirADAS/autonomous driving system incorporates state-of-the-artmodules for multi-class object detection, lane detection, depthestimation, and scene analysis. Optimized for complex roadconditions in cluttered urban scenes, this technology can accuratelyassess driving risks from pedestrian, bicycle, motorcycle, cars, andother types of obstacles. The solution being 100% vision-based, thismeans dramatic reduction in hardware cost, comparing to today’sLiDAR-based perception solutions. The company is headquartered inBeijing, China; with its research institute in San Diego, USA.

Before you ask, they are looking for talented minds and innovativehearts in CVPR community, with specializations in computer vision,machine learning, and GPU optimization - both in San Diego andBeijing.

10 CVPR ExpoCVPR Daily: Thursday

TuSimple

NB: Today (Thursday) is the lastday to vote for the People's Choiceaward. Make your vote count!

CVPR Expo 11CVPR Daily: Thursday

We noticed a cute demo at MorpXand decided to go and see it. TianLi was kind enough to tell us allabout it.

Tian Li: I am the founder ofMorpX. We are making a visionsensor for toy robots. The visionsensor uses machine learning todetect different objects like faces,bars, and lines using them tocontrol the toy robots.

CVPR Daily: What is novel in it?

Tian Li: The novelty is that it’smaking all of these algorithms runinside tight, small packagingembedded system. Also, its armcontact sensor with 72 megahertzfrequency.

CVPR Daily: What was particularlychallenging in doing this?

Tian Li: The challenge was makingall the right hardware andsoftware because the existinglibrary cannot fit into these kindof systems.

CVPR Daily: Does it make anydifference to work with toys?

Tian Li: There are lots of funstories because we are workingwith toys. When the algorithm isdefected, then everything fallsapart! So it’s fun.

CVPR Daily: So everybody plays inthe office?

Tian Li: Yea!

CVPR Daily: What is the next stepfor your application?

Tian Li: We are trying to pushthese products to classrooms forkids to have easy access to visionsensors to create some more funrobots. Our product is already inthe market and you can reserve itonline right now at our website.

“everybody playsin the office”

This black little robot has learned to hitdown only when he is shown a ball. Hedoes not react to any other object, nomatter how hard we try!

This is our last issue of CVPR dailyfor this year, so we have decidedto host two Women in computerVision. After Danna Gurari at page4, here is Angels Rates Borras.

CVPR Daily: Angels, what kind ofwork do you do?

Angels: I am doing a PhD atNortheastern University under thesupervision of Octavia Camps. Myfield of research is computervision. That’s why I’m here.

CVPR Daily: Why did you chooseto dedicate your career tocomputer vision?

Angels: I first studied electricalengineering back in Barcelona.Then I did an exchange atNortheastern University. I realizedthat I wanted to work withsomething related to artificialintelligence because I really enjoytechnology as well as science andmaths. I met Octavia and I likedher so much that I decided to jointhe lab.

CVPR Daily: Do you find it moredifficult to be a woman in thisfield?

Angels: In terms of studying andgetting opportunities, I don’t thinkso. I believe that if you’re good,they’re going to hire you. In termsof your team at work, there mightbe some differences between awoman and a man.

12 Women in Computer VisionCVPR Daily: Thursday

Angels Rate s Borra s

CVPR Daily: Can you tell usany stories about that?

Angels: I had a very funcollege experience. Back inBarcelona, you have toadapt yourself to work withmen because they workdifferently. Sometimes,they don’t know how towork together or how tocommunicate with eachother. Usually the womenare the ones that are incharge of organizingeverything and chasingthem to work…

CVPR Daily: Can you tell me thebest teaching that you receivedfrom Octavia?

Angels: Of course, her course atNortheastern. The way sheteaches is so lovely. She has a lotof patience. You can see in herface that she’s enjoying whatshe’s doing, and that’s the mostimportant thing. When sheteaches, she makes you feelwilling to learn and to listen toher. You always want to payattention. In general, she’s anamazing teacher.

CVPR Daily: Do you see yourselfteaching one day?

Angels: No, not for me.

CVPR Daily: What would you liketo achieve in this field?

Angels: I would like to learn a lot,that’s why I’m doing a PhD. Iwould like to publish someday. Iwould also like to work in a nicecompany.

CVPR Daily: Do you think thatyou’ll publish a paper one day?

Angels: Yes, I would like to publishsomething in computer vision, forsure.

CVPR Daily: What chance do youthink you have of achieving thisgoal?

Angels: For sure, I will do it!

Women in Computer Vision 13CVPR Daily: Thursday

“I wanted to work with something related to artificial intelligence

because I really enjoy technology as well as science and maths”

This project creates adaptive 3D sensors based on informationmaximization. Instead of treating 3D sensors as a black-box and askingwhat to do with the 3D data we get or where to move the sensor to learnmore about the scene, we are asking: if I already know something aboutthe environment, and I have a specific task in mind, how should thesensor operate?

The CSAIL / MIT team focused on structured light scanners, thatemit/project light patterns over time and acquire images. The question is:which pattern should the sensor project next? For example, in order toreconstruct a specific object that moves, you may not need to illuminatethe whole scene. If you just need to localize in a known environment orhelp a robot avoid obstacles, you may need radically different patternsthan if you want full scene reconstruction for augmented reality.

This research paper explores such options based on informationmaximization and sensor planning. Taking these concepts from decisiontheory and robotics, it is shown that with the right probabilistic model,they can be used inside the sensor as well.

The team chose a probabilistic model that incorporates the scene andsensor pose uncertainty, and yet allows to approximate the informationgain between the acquired images and subsets of variables in the scenesuch as the sensor pose or aspects of the geometry. The model allows todo so in a highly-parallel way, which, it is hoped, will make it useful forreal systems.

Adaptive 3D scanners, and the concepts shown are expected to result inmore efficient and accurate sensors that are better suited to the kind ofmultiple roles we expects robots and mobile devices will play in thefuture.

14 PresentationsCVPR Daily: Thursday

Guy Rosman - CSAIL/MIT