hello_text_object_detection

detect locatin of hello in the image

tensorflow object detection api

installation

data preparation

bounding box label generatd using LabelBox

LabelBox coordinates origin is differnet from that required by the script which convert label csv to TFRecord. use this jupyter notebook to

• change bounding box coordiantes origin from bottom left to upper left corner
• resize all pictures so that the maximum size is 300 x 500, while keeping the aspect ratio
• remove pictures with illogical bounding box coordinates, e.g. xmin > xmax, from training, but keep them in testing
• create train.csv and test.csv

convert train.csv and test.csv to train.record and test.record using this script

create hello_label_map.pbtxt

training

is done on google cloud ml engine by following this

files needed are

• ssd_mobilenet_v2_coco.config, which can be downloaded from here
  • change num_classes from 90 to 1, see here
• script to submit job to google cloud ml engine
  • tar pycocotools-2.0.tar.gz by following this to make eval job work on cloud

tried to train using faster_rcnn_resnet101_coco, but keep getting this error message - The replica master 0 exited with a non-zero status of 1. Termination reason: Error. - so turned to use ssd_mobilenet_v2_coco

model export

download mdoel checkpoint files to local and use this script to export the model

demo

can be found here