Pubg Real Time Player Detection Using TF2

Steps for training the model on google colab

follow the below link and perfrom step by step execution
"https://colab.research.google.com/drive/1tO3nGFzFX5nvARWqdr7m5bMUVM9W5NaW#scrollTo=vMNPnML5y9b7"
(mail: [email protected] to get the access for the colab link)

Steps for training the model on local machine

1. Create seperate virtual environment for TF2 packages installation "conda create -n (name_of_ur_choice) pip python=3.8" .
   Make sure you have anaconda installed , or visit "https://www.anaconda.com" to download the package .
   

2. Now activate the created virtual environment "conda activate (name_of_ur_choice)"
   

3. Now get the tensorflow2.x by running command "pip install --ignore-installed --upgrade tensorflow==2.2.0" .
   Check if the install is proper by running "python -c "import tensorflow as tf;print(tf.reduce_sum(tf.random.normal([1000, 1000])))"" .
   It should run without producing any error .
   

Note : This step is Optional . If you have GPU Nvidia GPU (GTX 650 or newer) then only go for step 4 or else train the model on CPU or colab

4. If you have GPU in your system then make sure you have "CUDA Toolkit v10.1" , "CuDNN 7.6.5" installed and integrated with TF2.x . I have attached the links for downloading the above 2 packages .
   CUDA Toolkit v10.1 : https://developer.nvidia.com/cuda-10.1-download-archive-update2?target_os=Windows&target_arch=x86_64&target_version=10&target_type=exenetwork
   CuDNN 7.6.5 : https://developer.nvidia.com/rdp/cudnn-download
   If the above downloads and integrations are successful then this command should run without any error :
   "python -c "import tensorflow as tf;print(tf.reduce_sum(tf.random.normal([1000, 1000])))""
   If it gives any error then kindly fix and try again .
   

5. Now we download the TensorFlow Model Garden "https://github.com/tensorflow/models" You can either git clone https://github.com/tensorflow/models it to the working folder
   Or just download the zip from the link and extract at the working folder
   

6. Now we install Protobuf "https://github.com/google/protobuf/releases" look for "protoc-3.12.3-win64.zip" and extract it in suitable location .
   Make sure you are setting up the environment variables as well .
   Run the command "protoc object_detection/protos/*.proto --python_out=." to check the installtion .
   It should run without any error , It might not return output but thats fine as long as it is not returning any error .
   

7. Now we install COCO API , Run the following commands .
   a) -> pip install cython
   Visual C++ 2015 build tools must be installed and on your path for getting package (b).
   if you don't have it get it from this link "https://go.microsoft.com/fwlink/?LinkId=691126"
   b)-> pip install git+https://github.com/philferriere/cocoapi.git#subdirectory=PythonAPI
   

8. Now we install Object Detection API. Now cd to "/models/research/" and run the following commands .
   -> cp object_detection/packages/tf2/setup.py .
   -> python -m pip install .
   Now to test the installations run "python object_detection/builders/model_builder_tf2_test.py" being inside "/models/research/" .
   It should return "Ran 20 tests in 68.510s OK (skipped=1)" type of output at the end .
   

(Delete both files "images/train.csv , images/test.csv" and "test.record , train.record" if you are building the model with your custom dataset)
9) Make sure you have dataset with proper images and xml files (pascal voc format) .
Split the train and test data , the folder structure should be like this
"images/train/"
"images/test/"

10. Run "xml_to_csv.py" to get the .csv files for both test and train (custom) dataset .
    

11. convert the test and train .csv files into the .record files by running following commands .
    a) !python generate_tfrecord.py --csv_input=images/train_labels.csv --image_dir=images/train --output_path=train.record
    b) !python generate_tfrecord.py --csv_input=images/test_labels.csv --image_dir=images/test --output_path=test.record
    It will create two .record files inside of the working directory .
    

12. Download the model which you want to train from "https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2_detection_zoo.md"
    Extract the file twice and put inside of the working directory .
    I have used "EfficientDet D0 512x512" model here and have also given the downloaded files . (feel free to use whatever model u want)
    

13. make sure u have the "labelmap.pbtxt" and "efficientdet_d0_coco17_tpu-32/pipeline.config" files parameters according to your dataset .
    I have commented out everything inside both the files , and it is pretty intitutive .
    

14. train the model via following command , it will craete "training/" folder and will store the checkpoints .
    "!python models/research/object_detection/model_main_tf2.py
    --pipeline_config_path={efficientdet_d0_coco17_tpu-32/pipeline.config}
    --model_dir={training/}
    --alsologtostderr
    --num_train_steps={number of steps you want model to train for}
    --sample_1_of_n_eval_examples=1
    --num_eval_steps={num_eval_steps}"
    

make sure that u have passed the paths correctly .

15. Once the training is completed , generate the inference graph of the final checkpoint created during the model training . "inference_graph/" folder would be created .
    "!python models/research/object_detection/exporter_main_v2.py
    --trained_checkpoint_dir {training/}
    --output_directory {inference_graph/}
    --pipeline_config_path {efficientdet_d0_coco17_tpu-32/pipeline.config}"
    You can find your model inside "inference_graph/saved_model/saved_model.pb"
    I have uploaded the folder here so that u can avoid the training of the model (but it's better to train it if u have better dataset than given one) .
    

16. Now run the "prediciton.py" file which will look for all the images present inside of the "images/test" directory and will give the results with the bounding boxes on it .

!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! HAPPY CODING !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
@ Author
Ayush Mishra

Results returned by the model

Results produced during analysis