dmbelo / Object_detection

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Bosch Small Traffic Lights Dataset

Installation : We need this API downloaded and installed, https://github.com/tensorflow/model Also all these will be need only to train model, on Carla we dont need these we will need just classify.

Refer this to complete the installation: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/installation.md

After installation,we need to follow these steps mentioned below:
(it is not mentioned but we need this for train.py)
1.setup.py build
2.setup.py install
3. windows:set PYTHONPATH=G:\CarND\ObjectDetection\models-master;G:\CarND\ObjectDetection\models-master\research\slim
4. linux : export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim

FILES to genererate TFRecord.record

  1. dataset_stats.py: gives all the classes and nos of images per class
    *Used in generatin label map pb.txt

  2. generate_tfrecord.py: This files reads yaml files and takes iamges and generates the TFrecord files. *the filename args, are passed in flags over here.
    flags.DEFINE_string('input_path', 'additional_train.yaml', 'Path to the yaml input')
    flags.DEFINE_string('output_path', 'train_TFRecord', 'Path to output TFRecord')

  3. object_detection_tutorial.ipynb is the script where we test our images.(Load the model and run test.)

Bosch data

https://hci.iwr.uni-heidelberg.de/node/6132/download/8fec5eefe8aea975e15ecda1eec6fc0e

  • We need to just take rgb images and not riib images
  • Test Data has no yaml
  • use generate_yaml.py generate yaml on test Data.

Config files

Need to edit ssd_mobilenet_v1_coco.config for the model.  
- Nos of classes 14.  
- Path of Test and Train record. I have provided full path, Linux user can give relative path.  

Data Dir

For training we need to work on Object detection setup.The dir struct looks like this:

+data  
    -label_map file  
    -train TFRecord file  
    -eval TFRecord file  
+models  
  + model  
    -pipeline config file  
    +train  
    +eval  

Training

For train.py use command :

    python train.py --logtostderr --pipeline_config_path=G:\CarND\ObjectDetection\models-master\research\object_detection \models\ssd_mobilenet_v1_coco_11_06_2017\ssd_mobilenet_v1_coco.config --train_dir=G:\CarND\ObjectDetection\models-master\research\object_detection\data

https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/running_locally.md

About


Languages

Language:Jupyter Notebook 98.8%Language:Python 1.2%