AdaptiveControl

From SeedWiki

Contents 1 Experiments with Learning Thresholds with NormalHedge 2 Experiments with Audio Tracking 3 Experiments with Tracking the Lorenz Attractor 4 Experiments in the UAI2010 Paper 5 Newer Simulations for Paper 6 collaboration with Tom Bewley's group 7 Writeups 8 Experiments with Tracking 8.1 Face Tracking 8.2 Tracking Synthetic data in 1D 8.2.1 Using Dynamic programming 8.2.2 Using Particle filter 9 Old Pages 10 Code

Experiments with Learning Thresholds with NormalHedge

• ThresholdsNormalHedge

Experiments with Audio Tracking

• TrackingAudio

Experiments with Tracking the Lorenz Attractor

• TrackingChaos

Experiments in the UAI2010 Paper

Experiment details, raw data, and Figures NHPFBayesMMExperiments

Newer Simulations for Paper

• Comparison between NH-PF (PF version of NH) and the Bayesian algorithmPFNHBayesExperiments
• Other Experiment Plans PFNHBayesPlan

collaboration with Tom Bewley's group

• TrackingChaos

Writeups

• Terminology For NormalBoost
• NormalHedge and Active Learning

Experiments with Tracking

Face Tracking

• yoav
  • with σ_vx = σ_vy = 0.5 link
  • with σ_vx = σ_vy = 1.0 link
• shibin
  • with σ_vx = σ_vy = 0.5: link
• mom and kid
  • with σ_vx = σ_vy = 0.5: mom kid
  • with σ_vx = σ_vy = 1.0: mom kid
• sunsern
  • with σ_vx = σ_vy = 0.5 link
  • with σ_vx = σ_vy = 1.0 link
• boyko and ben
  • with σ_vx = σ_vy = 0.5: boyko ben
  • with σ_vx = σ_vy = 1.0: boyko ben

--Yoavfreund 00:09, 8 March 2009 (UTC) Clearly, σ_vx = σ_vy = 1.0 is better! Another direction that I would like to explore is tracking an arbitrary patch. In other words, using as the score some measure of the fit btwn a patch of image taken around a selected point in a previous frame and the location of the particle in the current frame. I expect this to work pretty well assuming that you choose a good patch to track on, i.e. something like a corner, so that the patch changes significantly if you move it in any direction. This can be tested for pretty simply in the first frame. If we could implement this in hardware than we can potentially track tens of points at the same time, giving us the ability to track complex movements such as turning of the head, facial gestures, etc.

Direct Links to m4v files:

• Tracking Mom
• Tracking Kid

Tracking Using NormalHedge with a particle filter (MCMC) implementation

Tracking mom	Tracking boy

The results presented here are using a particle filter with 1000 particles, i.e. 1000 scored locations per frame.

THis does NOT need to include the fixed locations that are used to acquire the face initially and when we lose tracking. Losing tracking should be easy to detect and regaining the face in 1/3 second is very reasonable.

Tracking Synthetic data in 1D

Using Dynamic programming

• Some Experiments with Noise and NH prior AdaptiveControl/NoiseExpt
• Some Experiments with Noise and memory efficient version of NH Prior NoiseExptMem
• Some Experiments with New Variant of NormalHedge NHBoxExperiments

Using Particle filter

• Experiments with a 1-D Particle Filter type Implementation of Normal-HedgeAdaptiveControl/PFNoiseExpt
• Experiments with Particle Filters and Particle Filters with Normal Hedge PFNHExperiments

• Experiments with Particle Filters with Normal Hedge on a Signal/Noise object PFNHSNExperiments

• Comparison between NH-PF (PF version of NH) and the Bayesian algorithmPFNHBayesExperiments

Old Pages

Old adaptive-control twiki page[[1]]

Code

The MATLAB code is in Mercurial repository on seed.ucsd.edu. You can use the following command to obtain a copy of the code.

hg clone ssh://seed.ucsd.edu//data/hgroot/particle-code