Accurate Landmark Positioning at City Scales

-
The full paper can be found here

Introduction -  ALPS aims to use images from Google street view to analyze different views of an object to triangulate it's location(City Scale). The main focus of the Paper being finding( position)common landmarks like signposts, store fronts ,water hydrands etc automatically with high accuracy in a given geographical area
How it works: 
     
  1. Scours tools like street view to gather images
  2. Runs off the shelf object detector for finding the landmark
  3. Finds the position using least squares regression
Applications 
Some of the applications by the authors sound really interesting like use in self driving cars to get better position estimates or using to keep track of municipal assets in a city which are being done manually now. Location of hydrants is another great use which may improve the response time for the firefighters, I'm not too sure about how firefighters locate hydrants but feel that the data isn't publicly available. Personally i think that many GPS devices can warn car drivers if they go above  the speed limit. The best application i think would be navigation to visually impaired people delivered through wearables
 
Main Components:
Seed Location Generation: consists of Base Image retrieval (gathering the images from street view), Landmark Detection (detect and localize landmarks) and clustering(grouping of instances of landmarks that  are same)
Landmark Localization: consists of Adaptive image retrieval (drills down to find more images of the landmark ) and landmark positioning (to triangulate landmark's position )
Pros:
  • Sample at street view's resolution and sample lazily (more sampling only if the object is detected in the first sample). This two phase method helps to reduce the number of images that are sampled 
  • Location hints to narrow the search space
  • False Positives -  After identifying object in image, run it on a zoomed version of that image to reduce the false positives
  • False Negatives - If object is not detected due to occlusion or poor lighting , ALPS tries to detect object in neighboring images
  • Their tests indicate good coverage and accuracy. 

Cons:
  • ALPS is linked too much with google's street view at this point. As the authors mentions taking images from other sources also helps increase the coverage. The downside to this would be having to run the detector on that many more images
  • It takes a significant amount of time to run the object detection 
  • Even though they were able to lower the number of samples, It still takes time significant time to collect these samples

Comments

Post a Comment

Popular posts from this blog

Photo Tourism - Exploring Photo Collections in 3D

-
Image
Photo taken from here The Full Paper can be found here  and more information can be found  here Introduction: The authors create a novel photo explorer for browsing and organizing large collections of photographs by exploiting 3D geometry of the underlying scene. Their reason for coming up with the system is that the technology for browsing and organizing large amount of photos is old and will generally just show them as thumbnails and i also feel that this is true.  First Impression: At first I thought that some of the things that the system was trying to accomplish were not that hard to do like assigning a location or, were already done like scene visualization which i thought to be similar to Google's street view. Some features however did sound very interesting like like object based photo modeling which basically gives more similar images of the object  or of a particular item in the scene(In my experience Go...
Read more

Why start blogging

-
Image
Never did I ever think that I would start a blog, I am just not that person. So Why start blogging and what do I even blog about: I love technology, that is one of the reasons I decided to do my Master's in Computer Science and plan to start a career as a software developer soon. While I was doing my Master's I got an exposure to a few things that I really like some of them being Computer Vision - It has become mind boggling to me what the human eyesight is capable of and making a machine reach that level of perfection on such a staggering number of scenarios is simply put a challenging problem and one that I want to learn more about. Robotics - My interest for this field stems from the fact that I have absolutely loved cars ever since I was a kid, and ever since I found about autonomous cars, I developed an interest in robotics as well. I took a robotics course and understood and appreciated the struggle for providing a machine with autonomy. These are not the o...
Read more

Interactive 3D Modeling of Indoor Environments with a Consumer Depth Camera

-
The link for the paper can be found here Introduction The paper aims to develop a framework that will help reduce the complexities of 3D modelling so that it can be used by normal users. The system uses both color and depth for this, provides feedback and can tolerate human errors (alignment etc) and is capable of complete scene coverage.  Strengths Lowers the complexity for building 3D models  Makes it possible for non-experts to build dense models in real time on their laptops and gives feedback to the user with cues for alignment  Maps that are generated are metrically accurate (can be used measure dimensions ) The authors provide a video demonstration of the capabilities of their system and it seems quiet impressive you can check out the video here - https://www.youtube.com/watch?v=KJ6v37GypoE  , this video showcases user interaction (rewind and resume, loop closure detection ), 3D localization and using gestures in generated maps Compares be...
Read more