School-related news since arriving at Stanford in September 2003:

(7/16/2006) I have just released the first version of software developed to localize a vehicle based on various sensor data.  You may find the software on the Trajectory Smoother Software Homepage.  This has been released under the MIT license, which allows almost completely unrestricted use of the software.

(6/27/2006) I just returned from CVPR 2006, where a paper on some joint work I did with Steve Seitz, Brian Curless, Daniel Scharstein, and Rick Szeliski was presented by Steve and Brian.  The title of the paper is A Comparison and Evaluation of Multi-View Stereo Reconstruction Algorithms.  The related homepage for the project is the Multi-View Stereo Evaluation webpage.  The research page contains more details.

(10/17/2005) I attended the International Symposium on Robotics Research (ISRR) in San Francisco last week from 10/12 to 10/15.  It was a great conference.  I really like the single-track format and the papers were excellent.  It is also nice to see so much discussion after each session.  The sessions were organized to encourage this, with three or four 20-minute presentations, followed by 20 minutes of discussion.  This has two nice consequences: (1) it acts like a buffer, allowing the conference to stay on schedule despite some talks running over and (2) it gives the audience enough time to really probe the authors on their opinions about the broader field.  This was one of the most interesting parts about the conference because people asked some really difficult questions.

(10/17/2005) The Stanford Racing team won the DARPA Grand Challenge on 10/8/2005.  It was an exciting day down in the desert near Las Vegas.  The modified VW called Stanley finished the race in 6h54.  The next best times were 7h04 and 7h14, both from CMU.  Gray Team's vehicle finished in 7h30 and TerraMax finished in 12h51.  The course was 132 miles.  The $2M prize money will be given to the Stanford AI Lab.  All these details and more can be found at the official Grand Challenge website.  There was lots of suspense during the day.  The first-seated CMU vehicle called H1ghlander had a strong start and looked like it was widening the lead with the second-seated Stanley.  But a little over halfway through the race H1ghlander slowed.  It is not clear why this happened but several members of the Red Team have speculated that it was due to engine troubles.  Another possibility is raised by the fact that the gimble for the laser range finder on their vehicle appears to have failed at some point during the race.  When H1ghlander crossed the finish line the laser was pointed straight down.  I don't know if this means it was broken or if this is just the resting position, but it did appear damaged to me.  It is possible that their race software would have slowed the vehicle to a safer speed in the event of the gimble failure, or that without the laser data they ran into some obstacles.  In any case, Stanley caught up and was ready to pass well before the 100-mile mark.  But DARPA paused Stanley (stopping his clock, of course) to give space to H1ghlander until a suitable passing zone could be found.  We had a bit of a fright at that point because DARPA didn't announce the pause and we thought Stanley may have crashed.  We were relieved about 6 minutes later when he started moving again.  A suitable passing zone was found at mile 102 and Stanley had the lead position from then on.  At the end of the day it appeared that Stanley had won but we couldn't be sure because the third-seated Sandstorm (also from CMU) had also been paused several times and had started some unknown number of minutes after Stanley.  DARPA presumably knew the winner from reading the internal clocks, but they didn't announce it until the next day.  I had already flown back to Palo Alto and found the news on the website as soon as I arrived.  I relayed the news to the guys down in Primm, but they didn't announce it at the race site itself until the ceremony around noon.

(10/01/2005) The DARPA Grand Challenge is coming up soon (October 8th).  The research group of which I'm a member is fielding a team called the Stanford Racing Team.  By all accounts things are going well at the qualification event, which is going on as I type this (10/1/05).  The race itself is in just a week.  I'll be heading down to join the rest of the team on the 7th.  I worked on the code to generate the baseline trajectory given the sparse GPS waypoints from DARPA.

(10/2005) I'm back from the summer at MSR.  I had a great time working with Drew Steedly, Rick Szeliski, Steve Seitz, Brian Curless, and Daniel Scharstein.  The main outcome of the summer will be a paper (a work in progress) that is a survey of the field of multi-view stereo vision algorithms.  The webpage for the project will be online soon.  For an idea of what the project is trying to do for the multi-view community, check out Szeliski and Scharstein's similar page for stereo vision algorithms.  I've also decided to post of a few videos, and a few other materials.  These can all be found on the research page.  The other project, mentioned below was an effort at a multi-view vision algorithm of our own.  This will unfortunately remain as a project for another intern, as I was not able to finish it up.

(8/2005) Our 2005 NIPS submission has been accepted.  The paper is on the topic of using high-resolution color photographs of a scene to up-sample low-resolution range images of the same scene.  In this work we develop a multi-resolution Markov random field over the pixels and voxels.  The Gaussian likelihood potentials are selected based on the observation that discontinuities in range often occur at discontinuities in color.  These results show that using color photographs to infer 3D structure is an effective way to get more information out of today's range sensors.  Further details can be found on the research page.

(6/2005 - 9/2005) During the summer of 2005 I worked as a research intern with Rick Szeliski at Microsoft Research in Redmond, WA.  We worked on a technique for inferring 3D geometry from large sets of camera images take from known vantage points.  To be more precise, another part of the project, which I didn't work on, was to use a feature tracker and a bundle adjuster to find the camera's position from a simple video sequence.  Our method for inferring the 3D geometry is based on a probabilistic, generative model in which both the surface geometry and the color distribution are varied to maximize the a posteriori probability of the observed images.  The results of this project should appear soon on the research page.

(3/2005) Our 2005 ACM Transactions on Graphics Journal paper on Bayesian mesh smoothing and decimation was accepted for publication in the next issue (to be delivered to ToG by 10/05).  The mesh smoothing part of our method combines Gaussian sensor potentials with non-linear edge-preserving smoothness potentials to smooth out the noise in bumpy surface meshes generated from uncertain measurements.  The second part of the algorithm takes a already smoothed mesh and decimates it into polygonal patches of nearly-uniform surface normals.  These patches can be rendered much more quickly than the full mesh, while preserving much of the essential detail.  The paper describes the details of our algorithm and compares our results to competing methods.  Further details, including images and animations, may be found on the research page.

(12/2004) I gave a workshop presentation at the NIPS conference on bayesian mesh processing.  Further details, including several animations, may be found on the research page of this website.

(12/2004) I have officially earned my Master's degree, although the ceremony will not be until June '05. 

(11/2004) I passed the aero/astro qualifying exams.

(9/2004) I presented my first Stanford paper at the IROS conference in Sendai, Japan.  The presentation is also available.  This project was for Stanford's computer vision course, taught by Sebastian Thrun.  I worked on it with a Swedish foreign exchange student named Kjell Reuterswärd.  Also involved were (now Professor) James Davis, and his former co-worker R. Gupta at Honda Research in Mountain View.  The project was to use a space-time active stereo vision sensor on a mobile robot to create a 3D model of a room.  The details of the project are in the paper, and some movies that go along with the presentation may be found on the research page.

(8/2004) In early August I attended the SLAM Summer School in Toulouse, France, and presented a poster.

(6/2004 - 9/2004) I worked during the summer at Bosch Research in Palo Alto.  I worked under the supervision of Michael Breunig on work that eventually led to the ToG paper on Bayesian mesh processing.  Further details on the algorithm may be found on the research page.

Research and Education before Stanford: