How to config to merely show proper resulting windows?

[nguyenhu...@gmail.com]

I'm trying Face Detector of EBLearn. When I test EBlearn on nens.gif, it results in a lot of windows (below and above the 'threshold' variable in best_dir.conf file). Would you please help me to display merely windows whose confidences are larger than the 'threshold' variable in best_dir.conf?

OS: windows 7 (64bit)

Output image: http://www.mediafire.com/download/x91a4exsiu4hp2j/resulting_face.PNG

[Hoang Nguyen]

This is my best_dir.conf file

################################################################################

# META_TRAINER CONFIGURATION

# Note: variables starting with "meta_" are reserved for meta configuration

# required variables

meta_command = objrec_train

# optional variables

meta_name = face

meta_max_cpus = 8

#meta_output_dir = /home/sermanet/texieradata/face/out

meta_output_dir = /home/sermanet/humairadata/face/out

meta_email = m...@mail.com

meta_email_period = 1 # send an email with plots after every n iterations

meta_gnuplot_params ="set grid ytics;set ytics 5;set mytics 5;set grid mytics;"

################################################################################

# LOCAL PROGRAM CONFIGURATION

# directories ##################################################################

#ebl= ${HOME}/eblpierre/ # eblearn root

ebl=E:/Source_code/EBLearn

#root=/home/sermanet/humairadata/face/ds

#root2=${HOME}/eblearn/demos/face/trained/

#root2=C:\Users\pierre\eblearn\demos\obj\face\trained\

# # network ######################################################################

# net_type=cscsc

# net_ih=32

# net_iw=32

# net_c1h=5

# net_c1w=5

# net_s1h=2

# net_s1w=2

# net_c2h=5

# net_c2w=5

# net_s2h=2

# net_s2w=2

# net_full=100

# absnorm=1

# color=0

# mirror=0

# use_tanh = 1

# network ######################################################################

net = cscsc # net archictecture

color = 0 # use color or not

manual_load = 0 # manually load weights for individual modules

arch = ${arch_${net}} # global architecture

mirror = 0 # mirror instead of zero-padding (default)

nonlin = tanh # stdsig # non-linearity module

# layers

c0 = conv0,addc0,${nonlin}0,abs0,wstd0

s1 = subs1,addc1,${nonlin}1

c2 = conv2,addc2,${nonlin}2,abs2,wstd2

s3 = subs3,addc3,${nonlin}3

c4 = conv4,addc4,${nonlin}4

f5 = linear5,addc5,${nonlin}5

# machines

arch_cscsc = ${pp_detect},${c0},${s1},${c2},${s3},${c4}

arch_cscscf = ${pp_detect},${c0},${s1},${c2},${s3},${c4},${f5}

answer = class_answer

background_name = bg

# preprocessing modules

pp_y = rgb_to_y0

pp_yuv = rgb_to_yuv0

pp_yp = rgb_to_yp0

pp_ypuv = rgb_to_ypuv0

rgb_to_ypuv0_kernel = 7x7

rgb_to_ypuv0_mirror = 1

rgb_to_yp0_kernel = 7x7

rgb_to_yp0_mirror = 1

#resizepp0_pp = rgb_to_ypuv0

resizepp0_pp = rgb_to_yp0

#resizepp0_fovea = ${fovea}

resizepp0_zpad = 

pp_train = #rgb_to_ypuv0 #mschan0

pp_detect = resizepp0

# modules parameters

inputh = 32 # input's height

inputw = 32 # input's width

conv0_kernel = 5x5 # convolution kernel

conv0_stride = 1x1 # convolution stride

conv0_table = #${table0_color${color}} # convolution table (optional)

conv0_table_in = 1 # conv input max, used if table file not defined

conv0_table_out = ${table0_max} # features max, used if table file not defined

wstd0_kernel = 5x5 # normalization kernel

wstd0_gaussian_coeff = 0 # normalization kernel coeffcient

subs1_kernel = 2x2 # subsampling kernel

subs1_stride = ${subs1_kernel} # subsampling stride

addc1_weights = # weights to be loaded if manual_load = 1

conv2_kernel = 5x5 # convolution kernel

conv2_stride = 1x1 # convolution stride

conv2_table = ${table1} # convolution table (optional)

conv2_table_in = thickness # use current thickness as max table input

conv2_table_out = ${table1_max} # features max, used if table file not defined

conv2_weights = #${wroot1}${sp0}${sp1}_layer1_convolution_kernel.mat

addc2_weights = #${wroot1}${sp0}${sp1}_layer2_bias_bias.mat

diag2_weights = #${wroot1}${sp0}${sp1}_layer4_diag_coeff.mat

wstd2_kernel = 5x5 # normalization kernel

wstd2_gaussian_coeff = 0 # normalization kernel coeffcient

subs3_kernel = 2x2 # subsampling kernel

subs3_stride = ${subs3_kernel} # subsampling stride

addc3_weights = # weights to be loaded if manual_load = 1

conv4_kernel = 5x5 # convolution kernel

conv4_stride = 1x1 # convolution stride

conv4_table_in = thickness # use current thickness as max table input

conv4_table_out = noutputs # use number of classes as max table output

linear5_in = 10 # linear module input features size

linear5_out = noutputs # use number of classes as max table output

# tables #######################################################################

tblroot = ${ebl}/tools/data/tables/ # location of table files

# # conv00

# table00 = ${tblroot}/table_2_6_connect_6_fanin_0_density_0.5_uv0_u3_v6.mat

# conv0

table0_max = 6 #64 # full table output max (overridden if table file defined)

# table0_color0 = # no color, use table0_max for full table

# table0_color1 = ${tblroot}/table_3_32_connect_32_fanin_0_density_0.33_yuv0_y26_u29_v32.mat

# # conv1

# tbl=3

# table1_max = 64 # 64 96 128 256

# tbl1_mr1 = ${tblroot}/table_32_64_connect_1664_fanin_26_density_0.81_random.mat

# tbl1_mr2 = ${tblroot}/table_32_64_connect_1664_fanin_26_density_0.81_random.mat

# tbl1_2 = ${tblroot}/table_32_96_connect_2496_fanin_26_density_0.81_random.mat

# tbl1_mr1c = ${tblroot}/table_38_68_connect_2040_fanin_30_density_0.79_random.mat

# tbl1_mr2c = ${tblroot}/table_38_68_connect_2040_fanin_30_density_0.79_random.mat

#tbl_mrc = ${tbl1_3}

#tbl_mr = ${tbl1_1}

table1 = ${tblroot}/table_6_16_connect_60.mat

#table1 = ${tbl1_${net}}

# training #####################################################################

ds=1

val_size=500

dsname=all_mean32x32_ker7_bg_20100223.002945.face_conf00_eta_.00001

train=all_mean32x32_ker7_bg_20100223.002945.face_conf00_eta_.00001_train_500_1

val=all_mean32x32_ker7_bg_20100223.002945.face_conf00_eta_.00001_val_500_1

eta=.000005

iterations=150

# retraining ###################################################################

retrain=1

job_name_retraining=20100223.002945.face_conf00_eta_.00001

fp_name=_20100223.002945.face_conf00_eta_.00001

#retrain_weights=/home/sermanet/eblearn/demos/objrec/face/trained//20100223.002945.face_conf00_eta_.00001_net029.mat

retrain_weights=${root2}/trained//20100223.002945.face_conf00_eta_.00001_net029.mat

# preprocessing ################################################################

resize=mean

normalization_size=7

# tracking ####################################################################

mainsleep = 5

smooth_factor = 1.0

# display tracking

tracking_display = 0

# detection ####################################################################

#weights=${root2}/trained/20100227.175046.face_conf05_eta_.000005_retrain_1_net044.mat

classes=${root2}/trained/20100227.175046.face_conf05_eta_.000005_retrain_1_classes.mat

weights=E:/Source_code/EBLearn/demos/face/net00030.mat

#classes=${val_classes}

#threshold=0.1

gain=1

#input_dir = ${root2}/../../data/face/

input_dir = imgs

input_list = nens.gif

#show_conf = 1

input_height=-1#120#480

input_width=-1#160#640

#input_min=100

input_max=800

# multi-scaling type. 0: manually set each scale sizes, 1: manually set each

# scale step, 2: number of scales between min and max, 3: step factor between

# min and max, 4: 1 scale, the original image size.

scaling_type = 3

# scaling ratio between scales

scaling = 1.3

# scale factor of maximum resolution of the original resolution

max_scale = 1.0

# scale factor of minimum resolution of the original resolution

min_scale = 1.0

# number of detection threads

nthreads = 1

# randomize image input list (only works for 'directory' camera).

input_random = 0

# number of passes on the image input list (only works for 'directory' camera).

input_npasses = 1

# height factor to apply to bounding boxes

bbhfactor = 1

# width factor to apply to bounding boxes

bbwfactor = 1

# prune overlapping bounding boxes or not

pruning = 1

# minimum height ratio with smallest bbox to declare overlap

bbh_overlap = .67

# minimum width ratio with smallest bbox to declare overlap

bbw_overlap = 0

hzpad = .5 # vertical zero padding on each side

wzpad = .5 # horizontal zero padding on each side

mem_optimization = 1

max_bb_overlap = .7 # minimum ratio with smallest bbox to declare overlap

min_hcenter_dist = .3 # centers closer than this ratio over height cancel out

min_wcenter_dist = .3 # centers closer than this ratio over width cancel out

confidence_type = 2 # 0: sqrdist 1: single output 2: max other (recommended)

# detection display ############################################################

# output saving and display

save_detections=0

# save each classified frame and make a video out of it

save_video=1

# fps at which video should be constructed

save_video_fps=1

# if loaded a video and equal to 1, reuse video's fps

use_original_fps=0

# enable or disable display

display = 1

display_threads = 0 # each thread displays on its own

# only show classified input

minimal_display = 0

silent = 0 # minimize outputs to be printed

sync_outputs = 1 # synchronize output between threads

# display internal states of network

display_states = 0

display_min = -1.7

display_max = 1.7

display_in_min = 0 # input image min display range

display_in_max = 255 # input image max display range

display_bb_transparency = 0 #0.3

display_sleep = 0 #100000 # sleep in milliseconds after displaying

no_gui_quit = 1

ninternals=1

# demo display variables

queue1 = 0

qstep1 = 1

qheight1 = 5

qwidth1 = 2

queue2 = 0

qstep2 = 50

qheight2 = 5

qwidth2 = 5

# pre-camera variables (used before regular camera)

precamera = 0

precamdir = ${root2}/

# camera options: v4l2 opencv shmem video directory

camera = directory # v4l2 #opencv # directory

device = /dev/video0

file_pattern = ".*[.](png|jpg|jpeg|PNG|JPG|JPEG|bmp|BMP|ppm|PPM|pnm|PNM|pgm|PGM|gif|GIF)"

# limit of input video duration in seconds, 0 means no limit

input_video_max_duration=0

# step between input frames in seconds, 0 means no step

input_video_sstep=1500

job_name=20100227.175046.face_conf05_eta_.000005_retrain_1 # variable added by meta_trainer

#threshold = -1 #1.7 #.9

threshold = 0.9

threshold2 = ${threshold} #10

nms = 0

cluster_nms = 1

scaler_mode = 0s

bbox_max_overlap = .2 #.3 .4 .5 .6 .7 .8 .9

bbox_max_overlap2 = ${bbox_max_overlap}

bbox_max_center_dist = .9 #10

bbox_max_center_dist2 = ${bbox_max_center_dist} #.08 #10

bbox_max_wcenter_dist = .5 #10

bbox_max_wcenter_dist2 = ${bbox_max_wcenter_dist} #.08 #10

#display_sleep = 10000000 # sleep in milliseconds after displaying

scaling = 1.4 #2.2 1.9 1.6 1.3 1.1

# multi-scaling type. 0: manually set each scale sizes, 1: manually set each

# scale step, 2: number of scales between min and max, 3: step factor between

# min and max, 4: 1 scale, the original image size.

scaling_type = 3

max_scale = 1

#input_max = 288

show_extracted = 2

# DEBUGGING

#bbox_decision = 1

[Hoang Nguyen]

There are threshold and threshold2 variables in best_dir.conf. Are they different?

[Hoang Nguyen]

@the_minion: Please help me on this. Is the purpose of 'threshold' variable to display only detected regions having confidence larger than it? I did follow every step of your tutorials.

Thank  you.

[soumith]

Hoang,

Try increasing the "threshold" variable further, say for example 0.9999 or 0.99999

On Tue, Aug 13, 2013 at 11:59 PM, Hoang Nguyen <nguyenhu...@gmail.com> wrote:

@the_minion: Please help me on this. Is the purpose of 'threshold' variable to display only detected regions having confidence larger than it? I did follow every step of your tutorials.

Thank  you.

--
You received this message because you are subscribed to the Google Groups "eblearn" group.
To unsubscribe from this group and stop receiving emails from it, send an email to eblearn+u...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.
 
 

[Hoang Nguyen]

Hello the_minion,

Thanks for your response. I tried threshold = 0.99999, but the only resulting window shows just the same. There are very low-confidence faces which I underlined.

[Hoang Nguyen]

For the reason of readability, I show my conf file here:

threshold = 0.99999

[soumith]

Hmm, that looks weird. Let me check and get back to you. 

[Hoang Nguyen]

Thank you in advance.

[soumith]

Hoang,

Please add two additional variables:

pre_threshold = ${threshold}

post_threshold = ${threshold}

I'll try update the tutorial soon.

--

S

[Hoang Nguyen]

It really works. Thank you so much. You make my day.

What is the difference between pre_threshold and post_threshold btw?

Sincerely,

Hoang.

[soumith]

pre_threshold is the threshold to discard bounding boxes before nms (non-maximal suppression), and post_theshold is for after nms.

http://eblearn.sourceforge.net/detect.html#non-maximum_suppression_nms

[Hoang Nguyen]

It's so clear. Regular NMS looks more reliable than voting IMO.

Thank you so much.