- Articulated Swimming Creatures by Jie Tan et. al actually performed work I had planned for my thesis, but never did complete. They simulated various creatures using different fluid simulation methods, and found they get different results. This was the same thing I found, and generalised via PAL and used to generate control algorithms that can transfer to the real world for an underwater vehicle.
- Predator: A Visual Tracker that Learns from its Errors by Zdenek Kalal went viral. I think it was a bit hyped (thanks to a clever name), but the TLD algorithm is interesting, and there are some worthwhile items to note. First of all it is an open source visual tracker, and that there is a freely available
Google Tech Talk on TLD. In essence, Zdenek asserts that a tracker based on similarity of image patches will always lose its target and demonstrates trackers with learnt models from image data perform very well. (Similar to how models improve tracking in OpenTL). Zdenek et al. describe a way to generate more test sets using two experts (positive - objects must remain nearby, negative - only one true location per frame) and leverages boosting to get results. The total flow of operations are then: 1. Tracking and prediction, 2. Expert systems plus Detection, into Learning. 3. Re-Detection, using learnings.
- Paul Firth published a great introductory article on Physics Engines. Very handy. While your at it, Wolfgang Engel did a great post on some maths basics.
- Hierarchical Approximate Convex Decomposition of 3D Meshes (HACD) (HACD paper) is a method for representing objects as compound convex hulls (speeds up collision detection), that seems to outperform any other method I have seen.
- C2A by Min Tang et. al is a method for speeding up continuous collision detection with conservative advancement. A good paper, and source code is available.
- PAPPE is a Pascal Physics Engine.
- Barbara Solenthaler and Markus Gross published work on Two-Scale SPH Particle Simulation, an obvious idea (speedup simulations by mixing level of detail for SPH, they get about 2.2x speedups), someone had to do it. The videos are interesting nevertheless.
- On the topic of SPH, PhysXinfo has an article on the latests nVidia PhysX/Novodex/ETH research on fluid systems, of particular interest is Solid Simulation with Oriented Particles. Another interesting post is by Miles Macklin on fluid simulations.
- Intel released 'Colony' a crowd-simulation system, comes with source code.
- Lectures on machine learning with Matlab. Covers SVM, NN, EM, Bayesian methods, and image analysis.
- Ant colony optimization will be the next google AI challenge.
- An XML Unified Robot Description Format (URDF), again, like many other formats before it it has inherit limitations. COLLADA 1.5 Kinematics (library_kinematics_models) is probably a better choice.
- On the topic of XML, and other data interchange formats JSON cpp is a C++ library to handle exchanging JSON data.
- ACRA, the Australasian Conference on Robotics and Automation will take place in December in Melbourne.
- And just incase you've been hiding under a rock for the last while, the next big DARPA competition is the ARM manipulation program. Lots of interesting results are coming out of that project, but thats for another time.
Using Machine Learning To Cut Down Surgical Videos
10 minutes ago