version_note.txt

Standard model object recognition matlab code

Last updated September, 2006

To do?  Not critical.

1. Since C filters only work along the same features (no mixing in the 4th dimension), we can save memory and reuse the same filter for different features.  C filter is boring, because it is mostly ones.  We could use it as ones(xx,yy,1,1) on many features, instead of ones(xx,yy,1000,1000).  Hence, we can set the loop on w in do_sp2.c to repeat.

================================================
September, 2006
Ver5.2

1. Made the code image-size independent.  This will be useful to make the codes available for the machine-vision benchmarks (like Caltech Databases).

2. Made some of the repeated filters [].  In such cases, filters in the first scale will be looked up.  This should save some memory.

================================================
September, 2006
Ver5.1:

1. Added some preliminary error check aux_error_check.m, in which definitions in init_filter_def.m are checked.

2. Made aux_get_dummy_resp.m, so that no dummy run through the model is necessary in init_filt_oper.m.  As a result, the initial step of setting up the filter takes less time.

3. Made some more comments in sample_script_learn.m, warning against learning too few features.

================================================
June, 2006
Ver5.0: Almost ready to be released.

1. Moved most of the filter definitions to init_filter_def.m, so that pooling ranges, shifts, and number of features to keep (out of many learned features) can be defined for multiple layers from a single file.  However, some initial parameters (S1 parameters, such as number of orientations, Gabor filter sizes, etc) are still defined in filt_get_S1.m.  

2. Following the above implementation, most of the filter initializations are doe by the same set of functions (filt_get_S/C_layer.m).

3. Cleaned up learning steps, so that unused learning parameters are no longer defined (eg. learn_height).  Also fixed the frequent bugs, coming from ill-defined number of features.

4. Detached learning stage, so it is more cumbersome (ie. the user will have to do it in two steps: first, save responses upto some C layer, and second, run learn_filters_load_save to memorize patches.)  There is no longer an option to "learn" when invoking the software (however, this option was not very useful anyway).

5. Deleted several misc files (test_, util_, etc.).

6. The flow of model is re-defined, so that [S1 C1 S2b C2b S2 C2 S3 C3 VTU].  The allowed computations are:
   upto C1 only: [1 1]
   bypass route only: [1 1 1 1], [1 1 1]
   main route only: [1 1 0 0 1 ...]
   both routes together: [1 1 1 1 1 ...].
The output can be saved under one directory, under the subdirectory named "Layer?".  The saved outputs (usually C-layer units) will be used to run on classifiers.