BVLC / DPD

Deformable part descriptor project page

Deformable Part Descriptors (DPD)

This code accompanies the ICCV 2013 paper Deformable Part Descriptors for Fine-grained Recognition and Attribute Prediction.

===

User Configuration

[TODO: streamline these directory paths for minimal user setup]

in dpd_set_up.m:

scratchdir = /scratch  %for KDES features, DPD features, etc

if strcmp(database, 'bird')
    dataset_base = /path/to/CUB200-2011 %you edit this
elseif strcmp(database, 'cub200')
    dataset_base = /path/to/CUB200-2010 %you edit this
elseif strcmp(database, 'human')
    dataset_base = /path/to/berkeley-human-attributes-dataset %you edit this
end

===

Running DPD+DeCAF demos

Here's how to automatically classify CUB200-2011 birds using Deformable Part Descriptors with DeCAF convolutional features:

TODO: set scratch dir somewhere (env var?) TODO: add scratch directory argument: dpd_decaf(dpd_scratch)

#Option 1: run DPD+DeCAF in one line
./run_dpd_decaf.sh #this script does each step of DPD+DeCAF pipeline

#Option 2: run DPD+DeCAF steps one at a time
./extract_dpm_parts.sh #calls into ffld_dpm/build/ffld
./run_dpd_decaf_features.sh #DeCAF convnet features on DPM parts
matlab
>dpd_decaf; %weak pooling, SVM training, SVM classification

#dpd_decaf is hard-coded for this config: (weak pooling; CUB200_2011; trainAndTest). 
# not too hard to modify, though.

===

Running DPD+KDES demos

Here's how to automatically classify CUB200-2011 birds using Deformable Part Descriptors with Kernel Descriptors (KDES):

matlab
>run_dpd('bird', 0, 0); %runs all images through each DPD pipeline step, in batch mode

%arguments: run_dpd_kdes('class', 0=weakPooling 1=strongPooling, 0=trainAndTest 1=trainOnly);
%assume: DPM part bounding boxes are already extracted and located in a subdirectory of dpd_scratch

===

Reorganizing directories

```Shell

move: dpd_set_up -> dpd_kdes_set_up thirdparty/ffld_dpm ./ffld_dpm

new: dpd/dpd_decaf mv train_test/dpd_decaf.m dpd_decaf/dpd_decaf.m mv ./dpd_decaf_features.py dpd_decaf/dpd_decaf_features.py dpd/dpd_kdes mv train_test dpd_kdes common/{cksvd*, kdes_cholesky} -> dpd_kdes dpd/old mv dpd/test_realtime old mv dpd/test_realtime_cpp old mv thirdparty old/thirdparty

mv fgcomp old/fgcomp  #just a script to convert FGcomp labels to pascal format

add: dpd_decaf_set_up.sh #just point to dpd_scratch directory location