3D Photography Class Project

This paper presents an algorithm for breaking a model into approximately convex parts. This is good for us; approximate convexity, as opposed to real convexity, means both that there can be error tolerance in the decomposition and that small features, hopefully suitable for matching against, remain even in the decomposed parts.

Another paper on the same subject as the one above.

This is here for completeness - we are aware of other segmentation methods - but it's not robust enough for what we need.

Globally deformed superqaudrics might be great for a first pass filter of the data. They can be computed offline and searched fast. The downside is that they aren't sensitive to small local features - the only features we will have left after a convexity decomposition! On the other hand, we hope they can be used to quickly weed out 90% or so of the database, so that we can concentrate on more likely matches.

Using spherical harmonic functions, they (rotationally invariantly) decompose a shape into a symmetry descriptor, computed offline. Matching is then done in the space of these descriptors. This is pretty much the state of the art in shape descriptors right now. The lightfield descriptors paper claims an improvement over this method, but it's not applicable to us. Unfortunately, this may not be too applicable either. The essential feature that this method uses is a measurement of a shape's self-symmetry, which for a set of entirely convex objects may be fairly high. This might be good for us, and it might not be - it's still somewhat unclear.

Very simply, this is a standardized database of 3D objects to test a shape matching algorithm against. It's used in a lot of the papers here and we plan to use it as well.

This is an intriguing discussion about resizing models to deal with anisotropy. The caveat we must keep in mind is that if we use this (and I don't see why we wouldn't) then all of the parts being matched to a particular object should have similar (within some error tolerance) anisotropy compared to the parts matched to them in the database. This isn't a major limitation - we can match naively and just add in that constraint at the same time we check for parts being in the same un-decomposed model. It's just something to be aware of.

The shape descriptors described here are essentially predecessors of the Princeton stuff. They aren't as robust, but they also aren't as dependant on symmetry. This is not my favorite choice, but it may be the best set of shape descriptors we're going to find for the job.

My conclusions: This is a nice way of finding an object based on a 2D silhouette. However, it's fairly useless for us, since the individual parts after a decomposition will look very similar as silhouettes. It would also be a very slow for us, since we have to compare part by part. Their method of solving for rotations might be useful, but I can't see it working in our system.

This is a survey of several topics, but there is some good shape matching stuff in here. If we end up using a cocktail of weaker techniques, as I suspect we might, instead of settling on one or two good matching methods, we will likely draw from some of this material.

Good news: it seems to work. Bad news: the huge simplifications, particularly the pre-orientation of the model, means that it's impossible to seriously compare this to other work. Also, the segmentation was done in a way which left a lot of features in each part, while our segmentation method will produce much more "featureless" parts. Nevertheless, it's a good place to start, and there's a nice survey of part matching here.

This paper discusses fitting superquadrics to range image point clouds.

The cone primitive is a deformed cylinder, so this was useful.

A book on the same subject as the above papers.

The GraspIt! simulator can plan a grasp on an object if it has a primitive representation.