SSD using Inception

Download the following file and modify it:

cd /tf_files/SSD-Tensorflow/VOC2007

wget https://raw.githubusercontent.com/tensorflow/models/master/inception/inception/data/process_bounding_boxes.py

1. Modify "process_bounding_boxes.py" as in /tf_files/SSD-Tensorflow/MY_PIPELINE and copy to 
   /tf_files/SSD-Tensorflow/VOC2007/ directory
   
   Run it as: 
    python process_bounding_boxes.py trainval
    python process_bounding_boxes.py test
   
   Remember to have 2 levels of hierarchy to reach *.xml files
   
   Running the modified script generates all necessary bounding box and synset files necessary to 
   convert to tf_records format

Worked on TinyImagenet data from Stanford

#Run the following to re-structure the TinyImagenet train/ directory:
  cd /tf_files/SSD_TENSORFLOW/MY_PIPELINE

  python process_bounding_boxes_for_tinyImagenet.py /tf_files/SSD-Tensorflow/TINY_IMAGENET/tiny-imagenet-200/train

#Run the following to re-structure the TinyImagenet val/ directory:
 cd /tf_files/SSD_TENSORFLOW/MY_PIPELINE
 python process_bounding_boxes_forTIval.py /tf_files/SSD-Tensorflow/TINY_IMAGENET/tiny-imagenet-200/val

# Then copy the *_bboxes.txt file one directory level up and concatenate them together
cat train_bboxes.txt newVal_bboxes.txt > train_val_bboxes.txt

Now download this file:

wget https://raw.githubusercontent.com/tensorflow/models/master/inception/inception/data/build_imagenet_data.py

python build_imagenet_data.py \
  --train_directory=/tmp \
  --validation_directory=/tmp \
  --output_directory=/tmp \
  --train_shards=1024 \
  --validation_shards=128 \
  --num_threads=8 \
  --labels_file=voc_trainval_synset.txt \
  --imagenet_metadata_file=voc_trainval_synset_label.txt \
  --bounding_box_file=voc_trainval_bbox.txt


In our case do:
 mkdir /tmp/out_inception
 cd /tensorflow_models/models/inception/inception/data
 TRAIN_DIR=/tf_files/SSD-Tensorflow/TINY_IMAGENET/tiny-imagenet-200/train
 VAL_DIR=/tf_files/SSD-Tensorflow/TINY_IMAGENET/tiny-imagenet-200/newVal
 OUT_DIR=/tmp/out_inception

 python build_imagenet_data.py \
   --train_directory=${TRAIN_DIR} \
   --validation_directory=${VAL_DIR} \
   --output_directory=${OUT_DIR} \
   --train_shards=1024 \
   --validation_shards=128 \
   --num_threads=8 \
   --labels_file=/tf_files/SSD-Tensorflow/TINY_IMAGENET/tiny-imagenet-200/wnids.txt \
   --imagenet_metadata_file=/tf_files/SSD-Tensorflow/TINY_IMAGENET/tiny-imagenet-200/words.txt \
   --bounding_box_file=/tf_files/SSD-Tensorflow/TINY_IMAGENET/tiny-imagenet-200/train_val_bboxes.txt


# The TFRecords are generated in: /tmp/out_inception/

Some important links:

https://github.com/tensorflow/models/blob/master/slim/datasets/dataset_factory.py

https://github.com/tensorflow/models/tree/master/slim

https://github.com/tensorflow/models/blob/master/inception/README.md#getting-started

https://github.com/tensorflow/models/tree/master/inception

https://github.com/tensorflow/models/blob/4cc1fa0f58950b32532b5655aa46d3fb98bcd612/attention_ocr/python/data_provider_test.py