15-821/18-843: Mobile and Pervasive Computing

Project Descriptions and Mapping to Students

Fall 2017
Course Web page at http://www.cs.cmu.edu/~15-821




4
Just-in-Time Sign Translation
Anagh Singh, Vignesh Shankar
8(a)
Efficient Search of Edge-Sourced Data
Sandeep Agarwalla, Sindhu Simhadri
8(b)
Improving Search Efficiency in Edge Computing
Ziqiang Feng, Yanzhe Yang
9
ActiveMap Location-based  Services
Ticha Sethapakdi, Diana Zhang


4. Just-in-Time Sign Translation
(Mentor: Junjue Wang)

Travelers in a foreign language environment often have troubles navigating through surroundings. Sometimes failure to navigate can result in severe consequences, for example, going through wrong declaration lanes at Customs in airports. More often, it causes inconvenience and frustration, for instance, missing a highway exit. This project aims to perform just-in-time sign translation in places including airports, highway, and hotels, to help travelers navigate in a foreign language environment.

Leveraging a head-mounted smart glasses (e.g. Google Glass), the system translates phrases on signs as a user looks at them and displays the translation on the screen. The function is similar to Google and Translate, but the project should reduce end-to-end latency significantly leveraging cloudlets, since a user is likely to be walking or even driving as she passes through a sign. To build such a system, OCR is needed to convert images of text into text. Machine translation is also required to convert the text into another language. Fortunately, there are well established open-source projects for both OCR (tesseract-ocr) and machine translation (OpenNMT). The source and target language can be fixed to make the machine translation easier. The focus of this project is on overall system design and implementation, not on improving OCR nor machine translation.

8(a). Opportunistic Search of Edge-sourced Data
(Mentor: Jan Harkes)

Cloudlets receive multimedia data feeds from IoT sensors, video cameras, drones and other data sources that they are associated with.   Searching this data at Internet scale with relatively low result latency will become important in the future.   The goal of this project is to create a search framework for edge-sourced data that is based on the publish-subscribe paradigm.   Cloudlets can subscribe or poll for any appropriately scoped searches, run the search and   send results back to the cloud.

8(b). Improving Search Efficiency in Edge Computing
(Mentor: Jan Harkes)

Some edge computing nodes (e.g., on a small, autonomous drone)  may be under-powered, and incapable of running expensive algorithms such as those based on convolutional neural networks.   The cloud has the necessary resources, but shipping all the data from the edge to the cloud is wasteful of bandwidth.   However, the edge node could run a cheap "early-discard" algorithm that has a higher false positive rate than the best algorithm.  The goal of this project is to explore this tradeoff space and to demonstrate a search capability based on these principles.

9. ActiveMap Location-based Services
(Mentor:  Jason Hong)

The goal of ActiveMap is to make it easy for people to quickly create and deploy many kinds of location-based services. As an analogy, just as a web browser can read and render web pages, we want the ActiveMap to be able to read and render Maps. Example Maps might include a map of a location (e.g. a building floorplan), or content about where static or dynamic objects are (e.g. nearby Starbucks and discounts, real-time location of a person or members of a family, or location of buses as they move about). Systems issues include specifying the Map functionality and syntax (e.g. defining floorplans, defining locations inside, update rates, diffs from last read version, text snippets to
show), creating a GUI to display and interact with multiple Maps, and creating backend software for making these Maps easy to deploy on web servers.

Last updated 2017-09-18 by Satya