Code Checking by Version Cross-Validation
Note 1: In bin-chain version, the search order for finding the maximum in
the pursuit process is different from other versions, because of the putfront
operation. When there are ties (which can happen due to the local maximum
pooling), the result can be slightly different.
Note 2: The imresize function with the nearest option can be sensitive to
discretization error (e.g., 1e-16). Such an error can exist even if it should
be zero mathematically, and it can manifest itself in finding the nearest
neighbor in imresize.
(1)
old version vs
bin-chain version
check the basic learning code based on threshold
transformation. The two versions use different filtering functions.
(2)
bin-chain version vs
new basic version
check the basic learning code based on sigmoid
transformation.
(3)
new basic version vs
full version
check the basic learning code and detection code.
(4)
integral map version vs
slow version
Or
integral map version vs
slow version
Or
integral map version (double) vs
slow version (double)
check the code for computing window average (up to float precision).
(5)
integral map version vs
slow version
Or
integral map version vs
slow version and
another slow version (only differs in (xhere, yhere) part)
Or
integral map version (double) vs
slow version (double)
check the code for computing window average in nonaligned learning
(up to float precision).
(6)
single-scale version vs
multi-scale version
check the nonalign learning code (see Note 2)
(7)
past bin-chain version vs
recent single-scale version
check the nonalign learning code (see Note 1). The result is not as
pretty as that shown in Experiment 5. This is due to boundary handling
of local maximization. It may be better to force the boundary pixels to be
0 and do not perform local maximization on boundary.
See
zero boundary version.
(8)
Zhangzhang's comparison of different implemenations of adaboost
Matlab version and mex-C version. The search orders (for loops)
in maximization are different in these two versions, so if there are ties,
the results can be slightly different.
(9)
bin-chain version vs
new version
check the EM mixture clustering code
(10)
basic version vs
piece-wise version
check the code for learning from non-aligned images
Back to
active basis homepage