DPM Deformable Part Models release5新源码出来了
DPM Deformable Part Models release5
新代码出来了
What's changed since voc-release4?
voc-release5
---
The main highlights in this release:
* Weak-label structural SVM (wl-ssvm)
TheLearning code has been generalized to support both lsvm and
wl-ssvm.By default lsvm is used for training mixture models
andwl-ssvm is used for training the person grammar model from [1].
* Person grammar model [1]
Theperson grammar model from [1] is implemented in the person_grammar
directory.
* Optimization improvements
Thestochastic subgradient descent solver (learn.cc) has been replaced
byL-BFGS (using Mark Schmidt's minConf package). The learning and
featurevector cache code has been completely rewritten from scratch.
Notably,the feature vector cache is now stored in memory and not as
alarge file on disk. The memory requirements are slightly higher than
before,but your disk will thank you.
* Code cleanup, reorganization, and speedimprovements
Thegrammar API has been simplified and streamlined. The code has been
organizedinto subsystems. Training models is faster. For reference,
aPASCAL VOC 2007 bicycle model can be trained in 2 hours on a 4-year-old
quad-coremachine (using 'matlabpool open 8').
* Scale prior
Grammarproductions now have learnable "scale prior" parameters that
helpcalibrate detection scores at different scales. In conjunction with
this,the feature pyramid code supports adding an extra octave of HOG
featurescomputed on 2x2 pixel cells (off by default).
* Bug fixes
Severalbug from voc-release4.01 (mostly involving memory usage) have
beenfixed.
* New configuration system
Theglobals.m configuration file has been replaced by voc_config.m.
Thispermits overriding configurations by setting a global variable.
Suchconfiguration overriding makes it easy to test different config
optionson different machines from the same code directory