# AStar_Dual_Tree_HandPose
A fast NN hand-pose estimation system using dual-KD-trees.
Also includes implementations for various-SVM (inc. part models) and
tree based methods.

Please direct questions to [email protected]

Requirements
(1) OpenCV >= 3.0
(2) A C++11 compatible compiler
(3) Boost
(4) OpenNI 2.2 is needed to train/test with .oni files
(5) GLUT, OpenGL and LibHand are required to synthesize training data
    (this requirement can be disabled by editing local.cmake)
(6) OpenSSL, required for computing hashes for cache IDs. 

Compile w/ 
(1) cmake . 
(2) make -jX

Run as
./deformable_depth eval_model CFG=myconfig_file.cfg

# Configuration file options:
HEAP_PROF=FALSE # profile where the memory is being used...
JOINT_C=10 # for using SVMs
NUM_CPUS=20 # how many CPUs to use?
MAX_TRAIN_MINS=150 # cut off training at this point.
OBJ_DEPTH=20 # how deep is the bounding box
SKIN_FILTER=TRUE # use a high recall skin detection pre-processor
MANIFOLD_FN=ApxMin # which depths to consider?
NEG_SET=BIG # how many negatives to use when training? 
TEST_VIDEO=TRUE # test on the video sequences? 
# Cheat options for debugging
CHEAT_DEPTH=FALSE # give ground truth depth
CHEAT_HAND_BB=FALSE # give ground truth bounding box
# .985 is the highest we can go...
CHEAT_HAND_BB_THRESH=.8
CHEAT_HAND_ORI=FALSE # give ground truth orientation
#CHEAT_HAND_LR=RIGHT
CHEAT_ONLY_FREESPACE=FALSE
# DONT_USE_ARM=TRUE
OUT_DIR=out/
#OUT_DIR=automatic
# ApxNN, KinectPose, AONN, DeepYi, FLANN, NYUModel, Human, VolNN, Export, etc.
# Keskin
MODEL=ApxNN # use fast NN
WRITE_MODEL=FALSE
# SAVED_MODEL=/home/jsupanci/Dropbox/out/2014.09.01-ICL-NN25k20-SavedTree/model.yml
#SORT_NODES=
SORT_TO_DEPTH=1
ORTHOGRAPHIC=FALSE # tends to introduce artifacts which hurts performance.
# for training decision trees
TREE_SPLIT_SIZE=50
ENTROPY_TYPE=shannon

# Configure FLANN
FLANN_SAMPLES=50
FLANN_LSH_TABLE_NUMBER=15
FLANN_LSH_KEY_SIZE=15
FLANN_LSH_MULTI_PROBE_LEVEL=2

# configure the dataset
#DATASET=NYU
#DATASET=synth
DATASET=ICL
#DATASET=DIRECTORIES
#DATASET=EGOCENTRIC_SYNTH
#DATASET=test_videos
#DIRECTORIES=/home/grogez/Egocentric_Synth_Poser/
#DATASET=NYU
ICL_BASE=/home/jsupanci/workspace/data/ICL_HANDS2/
#ICL_BASE=/extra/titansc0/jsupanci/data/ICL_HANDS2/

#SYNTHETIC_DIRECTORY=/mnt/data/jsupanci/Synth-75000-WristVariation/
#SYNTHETIC_DIRECTORY=/home/jsupanci/data/synth-yi/
SYNTHETIC_DIRECTORY=/mnt/big/shared_data/Synth-75000-WristVariation/
#SYNTHETIC_DIRECTORY=/extra/titansc0/jsupanci/data/Synth-75000-WristVariation/
LEVEL_DB=/scratch/jsupanci/leveldb/
#LEVEL_DB=data/annotation-database

#DATASET=test_videos
#SYNTHETIC_DIRECTORY=./data/oracle_synth_armless_depth/
# setup the 500 exemplar model
# 3000  / 500
# 15000 / 5000
#NPOS=15000
#NPOS=25000
NPOS=2500
NNEG=0
#SYNTHETIC_DIRECTORY=data/2014.05.14-Synth50
#SYNTHETIC_DIRECTORY=/mnt/big/shared_data/Synth-75000-WristVariation/
#SYNTHETIC_DIRECTORY=/mnt/big/shared_data/Synth-75000-WristVariation/
IMPLICIT_IN_PLANE_ROTATION=TRUE
DATA_SOURCE_PRE_ROTATE=TRUE
TRAINING_INSERT_LR_FLIPS=FALSE
# kmeans, pyramid
ASTAR_ADD_BOUNDING=FALSE
ASTAR_CLUSTERING_ALG=kmeans
DEBUG_ASTAR_VIDEO=FALSE
NN_XTEMPL_CACHE_SIZE=50000
# no pruning with 5000000, 500000
NN_SMA_OPENED_LIMIT=50000
NN_ADMISSIBILITY=1
NN_PYRAMID_SHARPNESS=.5
SEGMENTATION=FALSE
NN_LOG_TREE=FALSE

# Analysis Section - universal parameters
POST_POSE_CONVEXITY=FALSE 
POST_POSE_VORONOI=TRUE    # this one helps
INTERPOLATE_TRACK=FALSE # tries to make tracker evaluation more fair.
INTERPOLATE_PARTS_IK=FALSE
# factor by 2.5 to 4.
FINGER_DIST_THRESH=4
SCORE_SKIP_LEFT=FALSE
# Right dominstates in egocenric
SCORE_SKIP_RIGHT=FALSE

# Synthesis parameters
SYNTH_CLUSTERS=5
SYNTH_EX_PER_CLUSTER=25000
SYNTH_PERTURB_R_MIN=20
SYNTH_PERTURB_R_MAX=120
SYNTH_FINGER_AREA_FILTER=FALSE