Disney Developing Re-Identification Analytics

By Brian Karas, Published Oct 27, 2017, 09:27am EDT

When most people think of Disney, they probably think of cartoon characters or maybe innovative theme park rides and RFID bracelets. But Disney researchers have also developed an person-tracking analytic they claim can significantly outperform state-of-the-art approaches. [link no longer available]

We analyze Disney's approach, using dynamically trained convolutional neural networks to find appearances of the same person across multiple surveillance cameras.

Multi-Camera ****** ********

******'* ***** ***** ******* ** ************ ** "**-**************", the ******** ****** ******** terminology *** ******** *********** of a ****** ** ******** images. *** ***** ***** does *** ************* ** the ********* ** ******** for *** ** **** or ******** *****, ****** as-described, ** ***** ****** be **** ** ****** scenario, **** ****** ********.

Applications *** ******

****** ***** ***** ********* an "**********", ***** ** effectively * ****-**** *****, in ***** ** **** things "****-****", **** **** to ******* ****** ********, without *** ******* ********** of ***** **. ** a ****** ****** * disruption, ********** **** ****** from ******* ********** *********** (and ***** *** ********** for ***** *********) ** a **** ********. ***** able ** ****** * person ** **** **** from ******** *******, ******* of ******* ******** ******** guards **** *** ****, would ** ** *********** benefit. ****, **** ********** could **** ****** ******* their ******** ******* ** customers, *** ****-**** **********.

Overview ** ****** **-************** *********

****** ** ******** **** breaking ** ***** ** a ****** **** ******* sections (***** **** **** "patches"), *** ***** **** ***** to ***** ************ ***** algorithms ** * **** neural ******* ********* *** probability ** * *****. This ** *** ** looking *** ************ ** individual ******* ** *** image, ******* ** ******* only ** *** ***** image ** *** ******. If * ***** ***** is ***** ** ****** patches, *** ***** ** deemed ** ** ** the **** ****** ** the ******/******** *****. **** improves *********** **** *** person ** *** ****** seen **** *** **** angle, ****, ** ********/********** across ******** *******.

****** ******* **** ******** over *********** ******* ** relying ** ***** ** texture **** ******* **** data ** *** ****** reliably ******* (**** ** when *** ***** ** the ****** ** **** a **** ******** **** of ** **********).

Specifics ** "*****" ********

***** ******** ************* ****** ******** are ******* **** ********** of * ****** *****, in * ****-******** ********. The ***** (***** *) is ******* ***** * full-body ********** ** *** person, *** ****** (***** 1) ***** *** **** image, *** ****** **** horizontal ********, *** **** the ***** (***** *) slicing **** ** *** horizontal ***** **** ****/***** halves. **** ********** ***** ** ******** to ** * *****.

**** ********* ** ***** for *** **** ** patch ** **** ** fed. ****, *** *** Level * (****** *****) portion, **** ** *** 4 ******/******* ** *** individually ** *** **** algorithm ** ******* * separate ********. *** ******* **** distinct ***********, ** *** Level * ********* ******* at *** *** ***** ** images **** **** ** the *** ***** ** ***** images *** ********** ********.

* ********* **** ****** each ***** **** ******** sub-patches *** ******* ******** and deep ******* **********. **** stage ******* ** **** the ***-******* **** ******* of ******** ** *** image ******* ****, ***** the ***** ***** **** not ******* *********** ******.

 

*******, * *********** ****** learning ******* ** ******* out ** * ***** and ***** ***** ** the *******.

****** ****** ** **** 3-step ******** ** * Deep ******* ******* (***) method.

*********** **-************** **********

******** ** ******* ********** are ********* ** ******'* paper. ***** ******* **** of * ********** ****** of ****** ** * given ****** (***** ***** ideally ** ********* ** distinct ***** ******** ** the **** ******) ** train * ******* ****** network, *** ****** ** use ******* ******** ** "people" **** ****-****** ******* from ********** ******** *** differentiators ** * ***** person:

Claimed ******** ** ***** ********

** *** ********** ** the *****, ****** ****** that ** ********* *********** learning ********** ***** *** images *** ******** ** determine ********* ** ********* between **** (************** ******), **** *** *****-***** image ******** ********, **** can ******* ******** ******* for ******* ******* *********** of * ****** ****** multiple *******.

 

VIPeR ******* ****

** ******* ***** *** re-identification ****** ** *********** methods, ****** *********** **** the ***** *******. ***** ****** ** "Viewpoint ********* ********** ***********", which ** * ******* of ****** ** ****** captured **** ************ ******* having ********** ** ********/*******, poses, ******, ***. ** represent ******* **********. ******** of ***** ****** *** shown *****:

Performance ************ ** *********** *******

** ***** ***** **** to ******* ***** ******** to *********** **-************** *******, Disney ******** * *********** increase ** *********** **********.

*** **** **** ** the ***** ********** ***** the **** ******* ******* approach ** *** ***** 2 ******* (***** *** person's ***** ** ****** into * *******), **** overlapping ***-******* **** ********. The ***** **** ** using *** *** ********, but **** * *********** of * *********** ******* (color **********, ***/**********/***** ****, and ***** ***********), *** the *** **** ** the *** ******** ***** the ***** * ******* with ** *********** ****. The ***** *** **** lines *** *********** "**** crafted" ********** ***** ***** 2 ******* *** *** color ****:

 

********** **** ****** **** Disney ******** *** *** approach ****** **** * significant *********** ** **-************** accuracy, *** *** *** best ******* ***** ***** 2 *********** *******, ****** DSP ******* ** *********** input **** (***** ******) also ****** ***** ************ in **-**************.

Disney ** *******

****** ******** ******* ** IPVM ** ******* ******* of **** ********.

Impact ** ************ ********

******** ************* ***** **** ** sell ******, *** ***** large *********, ***** *** versions ** **-************** *********. **********************, *** ***** ********* companies **** *** ******** heavily ** ******** **** ************ for ***** **** ******. However, ************ ****** ********, *** ***** **********, are ****** ** ****** to ******* ********** ********* ** a ***** **** ** developers. ** ****, **** allows ************* **** ****** to ***** ***** *** advanced ********* *********, ******* of ********** **** *********** suppliers.

***** ** ** ******** that ***** ***** ************ will **** ** ******* their *** **-************** *********, or ******* *********, *** possibly **** **** ***** do ** ***** ** harder *** ************* ** charge ******* ******. **** may **** ** **** affordable *********, ****** ***** technologies ********* ** * wider ******* ** *****. 

Comments (4)

Dear Brian, thank you for this article. I think that one of the most interesting parts of the article is the last figure. We may consider it as a current state of research in the field of reidentification. RANK 1 below 50% (and RANK 15 near 90%) may be interpreted as an indicator that the task of reidentification is not reasonably solved even on a database like VIPeR (it contains only 632 persons). So I suggest that on real sites (with thouthands of different persons per a day and hungreds of cameras) the method will not work feasible. Also it is interesting what about CPU (or GPU) consumption.

 

AvigilonIronYunQognify, and other analytics companies have all invested heavily in building such applications

After this article from Disney Research it's interesting how these analytics work on real sites. Does someone have experience?

Avigilon would not cope well with the crowds.  I suspect other analytics would be much the same.

Brian, thanks for a very interesting article. Interesting that Disney would "go it alone" instead of partner for technology. Makes me wonder if that reflects their desire to keep the details of this project as contained as possible (not supported by letting a technical paper), or if they just couldn't find suitable commercial technology. 

Skip - hard to say since they did not want to participate in an interview, however given Disney's overall approach to technology/innovation, I did not find this overly surprising.

Read this IPVM report for free.

This article is part of IPVM's 6,738 reports, 909 tests and is only available to members. To get a one-time preview of our work, enter your work email to access the full article.

Already a member? Login here | Join now
Loading Related Reports