image feature file format and mirflickr-sift-fastkmeans-1000000-new.voc file format.

[Anonymous]

I want to use the FLANN instead of QuantiserTask class for visual word assignment. I want to know how the data is organized in image feature file xxx.jpg.fv and mirflickr-sift-fastkmeans-1000000-new.voc so that I can read them by C++. Any suggestion?

[Jonathon Hare]

Note all binary numbers are written in Java format (i.e. Big Endian) - if you want to read them in C++ on a little endian machine (i.e. any PC with a x86/64 type processor), you'll need to swap the byte order. The FV files are lists of local features, with the following header structure:

KPT<numfeatures|int><featureLength|int><features...>

The features are actually dependent on the type of local feature extracted... For DoG/SIFT each one looks like this:

<x|float><y|float><scale|float><orientation|float><f_1|byte>...<f_featureLength|byte>

The quantiser file format depends on what clustering mode you chose and what precision (byte or int). Assuming you chose KMeans or Fast-KMeans with byte precision (the default), the .voc file has the following format:

BCen<numClusters|int><featureLength|int><centroid_1_dim_1|byte>...<centroid_1_dim_featureLength|byte><centroid_2_dim_1|byte>...<centroid_2_dim_featureLength|byte>...

I'd be interested to know whether you find FLANN any faster... In theory there shouldn't be much difference as imageterrier is using the same fundamental algorithms for cluster assignment.