sofwerx / tensorflow-gun-detection

tensorflow-gun-detection

Scores

Model	Mobile SSD	Faster RCNN
Test Loss	1.5	0.1016
Trained Rounds	5114	7167

Instructions

• Uncompress the train & test images

  • Images inside images are compressed into train.tar.gz and test.tar.gz folders.
  • Navigate to images folder and type "bash uncompress.sh" in your terminal.

• Create csv records for the xml files

  • Navigate to images and run xml_to_csv.py file

• Create tfrecords

  • Navigate to images and see the generate_tfrecords.py script for instructions.

• Once you generate xml files and tfrecords, the files should be available under images/data

• Create a folder called training in the root folder.

• If you want to do transfer learning pull the model & config files from tensorflow zoo and put the in the root folder.

• Move the config file to training/ folder and change the config file to match the paths.

• To train the model:

  • python train.py --logstderr --train_dir=training/ --pipeline_config_path=training/ssd_mobilenet_v1_pets.config

• Export inference graphs

• Paste the graph folders to outputs/ folder

• Count the no.of bounding boxes

  • python image_detection.py

utils/image_resizer.py

• usage
  • python image_resizer.py -input=input_folder_images -output=output_folder -height=800 -width=600

Notes:

• To compress the images which are inside train & test folders inside data/images/processed

  • tar -cvzf train.tar.gz train
  • tar -cvzf test.tar.gz test

• Rectlabel App in Mac / LabelImg open source:

  • The bounding boxes around the images were created using Rectlabel tool available for MAC.
  • We can also use Labelimg open source tool for this task.

References:

• Modified version of xml_to_csv.py from racoon github repo.
• generate_tfrecord.py from racoon github repo.

tensorflow-gun-detection's Issues

Images dataset

Could you please share the link to training images for the dataset?

Typo at the print statement in line 38, in the file "xml_to_csv.py "

Typo:
print('Successfully converted xml to csv for, ' directory)

It should be:
print('Successfully converted xml to csv for ', directory)

Export inference graphs

Available in outputs folder.

Wrong detection

Hey there,
After the training (around 3k steps) it detects almost everything as a gun.
Example photos:

Total loss chart:

Should I train it more or I've done some mistakes in configuration?

Dependency problem

Running "generate_tfrecord.py" :

Traceback (most recent call last):
  File "images/generate_tfrecord.py", line 20, in <module>
    from object_detection.utils import dataset_util
ImportError: No module named object_detection.utils

Where is the object_detection file ?
Thanks.

mAP score for ssd mobilenet

Once training + prediction is completed. Find out how much mAP the model is giving.

DuplicateFlagError coming in tfrecord.py file

Cannot find train.py

There is no train.py in the git repo

No train.py in project

python train.py --logstderr --train_dir=training/ --pipeline_config_path=training/ssd_mobilenet_v1_pets.config

But there is no train.py in this project.

Bounding boxes count for both ssd mobile + faster rcnn inception v2