Pursuing Atomic VideoWords by Information Projection
The running process is introduced as follows:
There are some directory pathes that are needed to set by yourself according to organization of your data.
In addition, the 'lambdaTable_Qb' is implemented offline to establish a look-up table based on a number of natural video bricks for the parameters computing.
First, run the 'ExtrData4Sketch' and 'ExtrData4Texture' to extract the headers of all bricks (indexing each brick in the video set) and filter the video sequences with large and small scale Gabor filters respectively. For each video sequence, the cost in time of 'ExtrData4Sketch' is about half-hour (Core2 Duo CPU 2.66GHz with 2GB RAM) and in space is about 1GB. The costs of 'ExtrData4Texture' are about halfhour and 200Mb.
Then, run 'Preprocessing' to orgnize all the bricks in a efficient manner for the following process.
First, run 'OverPartition' to group all the bricks of each video sequence into a number of atomic clusters. For each video sequence, this process needs about halfhour. And the number of atomic clusters of each video sequence is about 2500.
Then, run 'DataPrepare' to combine all the atomic clusters of all videos to form a whole matrix of headers of leaves for the following clustering step. In this step, we also compute and save the mean of all bricks' responses of small scale Gabor and the sum of all bricks's responses of large Gabor in each leaf node for the convenience of following process. This cost about twenty minutes and 800Mb.
Run 'InitClustering_Stage2' to group all the atomic clusters into a number of clusters (e.g., K = 8000 clusters) preserving three nearest cluster centers for each atomic cluster. So there are 3K clusters in total. This step will need several days.
Words Modeling and 5. Words Pursuit
These two steps are implemented iteratively.
First, run 'model_pursuit_each' to learn the model of each cluster (or words) and compute the corresponding information gain.
Then, select the word with maximum information gain and re-compute the models of the words that share samples with the selected words.
This step also needs several days (about 3 days).
152754 visistors since August 2009. (C) Haifeng GONG, 2008-2012.