Skip to content

Latest commit

 

History

History
142 lines (76 loc) · 4.96 KB

README.md

File metadata and controls

142 lines (76 loc) · 4.96 KB
Raw

Introduction: This Model has been initially build and trained on local windows 10 machine equipped with 12GB RAM (No GPU and TPU)and will be trained later on TPU via Colab

Model has been trained to detect the following objects from street:

  • lane
  • pedestrian
  • sidewalk
  • vehicle

Street Vieews


Local Environment Specifications

  • Windows 10 Enterprise
  • python version(3.8.8)
  • TensorFlow version(2.4.0) (Facing issues with 2.4.1)

prerequisites

  • Installing Tensorflow object Detections API's you can refer to below link Installation Guide
  • Install wget on your machine to clone github repository

Note: As part of this installation many other dependencies need to be installed as well, please make sure to go through all installation steps, they are bit long and may you would face some challenges

Clone the Github repository

type thr line of code in anaconda prompt

git clone https://github.com/halhwadi/Instance-Segmentation---Tensorflow-Object-Detection-API-s.git


Downloading Images

street images have been downloaded by using google map street view and stored under this path (images/resize) only 30 images have been downloaded to avoid overloading our modest machine


Resizeing

IF your workstation is not equipped with high RAM its highly recommended to resize the images Note: In pipleline.config file the size in image resizer should be match with resized images you use in this step

You can resize the images in anaconda prompt by running the following line of code(make sure your images are available under this path (images/resize)...Width height

resize.py 400 300 images/resize/


Annotations

In this tutorial we will use Labelme tools, you can install by running (conda install -c conda-forge labelme) Once its installed please type labelme on Anaconda prompt to open it

Note: please avoid using circle as we faced issue while running create_coco_tf_record script to create tfrecord, this script throws error when number of annontation points is 4

Once annotation is done, please make sure to split images between train and test by moving them to 'images\train' and 'images\test' folders

Note: make sure to move the images and their corresponding json files as well


Creating COCO files

We need to create the json files, one for training images and the other one for testing images, In order to create these files "labels.txt" files need to be generated, its txt file and you can create it manually ( Note the ordering of classes in labels.txt file should be the same oredering in labelme desktop screen, you can review the labels.txt file for more details )

type the following commands in Anaconda prompt:

python labelme2coco.py images\train train --labels labels.txt

python labelme2coco.py images\test test --labels labels.txt

Sample of Annotated Images


Creating Label map:

Refer to tutorial notebook for more details


Creating tfrecords: running the below command in Anaconda prompt:

python create_coco_tf_record.py --logtostderr --train_image_dir=images/train --test_image_dir=images/test --train_annotations_file=train/annotations.json --test_annotations_file=test/annotations.json --output_dir=tfrecords --include_masks


Downloading the mask_rcnn model from Tensorflow model zoo:

Refer to tutorial notebook for more details


Preparing the Configuration file pipeline.config

Its recommend to use the same pipeline config file under model folder because many changes have been applied to this file to facilitate the training on windows machine with modest RAM (Example: number of batches set to 1), you can modify the number of classes and keep the rest of details as is, otherwise you can use the master file from Tensorflow zoo and apply the required changes by your self(Refer to tutorial notebook for more details)


Training

Running the below line in Anaconda prompt:

python model_main_tf2.py --model_dir=model --pipeline_config_path=model\pipeline.config --num_train_steps=3000


running in Anaconda prompt:

the below line of code: tensorboard --logdir=model\train


Inference:

Refer to tutorial notebook for more details

Author:

Husam Alhwadi
[email protected]