Face detection using my own dataset

[PiyalGeorge]

Hi, I'm new to machine learning, caffe, LMDB stuffs. Currently i'm trying to create a caffemodel using mobilenet ssd, using my own dataset. I'm using OpenImage's face dataset (Eg: open-image-dataset). When i extracted this dataset, i'm getting CSV files that has data in it, unlike the Annotations(xml files, etc). Anyway i downloaded the images. Now how should i create the LMDB with theses images? I want to detect person and faces from video. During LMDB we use some labelmap.prototxt file(actually what does this file do?) . how will we create this file for so many images?
Also incase if i have created this LMDB, how will i train a pretrained mobilenet ssd model with these LMDB to get the right detections results?

Please bro @chuanqi305 , help me. I have been looking for this for days. but still not getting the right answer

[chuanqi305]

@PiyalGeorge Firstly， you should convert the labels to voc format like this:

VOC2007
000001.jpg

The VOC2007 Database
PASCAL VOC2007
flickr
341012865


Fried Camels
Jinky the Fruit Bat


353
500
3

0

dog
Left
1
0

48
240
195
371



person
Left
1
0

8
12
352
498

And then generate label list file train.txt , val.txt and test.txt.
The labelmap.prototxt is used to convert the class name in xml to id.

[PiyalGeorge]

First of all i thank you @chuanqi305 for coming to help me.
I didn't get the use of that much of text files . @chuanqi305 this is what i have done:- from the Openimage's CSV dataset in above link, i have created the xml file for each image and stored in another folder. I downloaded all images to a folder. then i created a text file(one text file only) with <path_to_image>SPACE<path_to_its_xml_file> information.

The xml file looks like this:- 0a0c1f65f19859c0.xml ->

<annotation>
   	<folder>COCO</folder>
  	<filename>0a0c1f65f19859c0.jpeg</filename>
  	<source>
  		<database>The VOC2007 Database</database>
  		<annotation>PASCAL VOC2007</annotation>
  		<image>flickr</image>
  	</source>
  	<size>
  		<width>512</width>
  		<height>512</height>
  		<depth>3</depth>
  	</size>
  	<segmented>1</segmented>
  	<object>
  		<name>face</name>
  		<pose>Unspecified</pose>
  		<truncated>0</truncated>
  		<difficult>0</difficult>
  		<bndbox>
  			<xmin>216</xmin>
  			<ymin>0</ymin>
  			<xmax>278</xmax>
  			<ymax>64</ymax>
  		</bndbox>
  	</object>
        <object>
  		<name>person</name>
  		<pose>Unspecified</pose>
  		<truncated>0</truncated>
  		<difficult>0</difficult>
  		<bndbox>
  			<xmin>210</xmin>
  			<ymin>0</ymin>
  			<xmax>434</xmax>
  			<ymax>512</ymax>
  		</bndbox>
  	</object>
</annotation>

theni created a labelmap.prototxt on my own with following entries only:
item {
name: "none_of_the_above"
label: 0
display_name: "background"
}
item {
name: "person"
label: 1
display_name: "person"
}
item {
name: "face"
label: 2
display_name: "face"
}

in
So inputs now i have are - 1) Images 2)xml-files 3)text file that map image and xml file 4) labelmap.prototxt.
then i gave this command to create the LMDB:-

/caffe/build/tools/convert_annoset --anno_type=detection --label_type=xml --label_map_file=/home/Documents/labelmap_voc.prototxt --check_label=True --min_dim=0 --max_dim=0 --resize_height=0 --resize_width=0 --backend=lmdb --shuffle=False --check_size=False --encode_type=jpeg --encoded=True --gray=False /home/PATH_TO_DATA/ /home/Documents/mobile-ssd/text.txt /home/lmdb/VOC0712_test_lmdb

Here VOC0712_test_lmdb is the name of lmdb to be created.

But during these procedures where do i encounter train.txt , val.txt and test.txt ?
Also i dont know whether i'm doing things right. Please correct me bro @chuanqi305 if i'm going wrong. I will be eagerly waiting for your reply