https://www.nytimes.com/2018/03/15/nyregion/bike-lane-blocked-new-york.html

OurCamera

Read About Project Here

This project uses Google’s TensorFlow Machine learning package to identify and categorize types of vehicles on NYC streets.

The data images are downloaded from NYC DOT cameras.

The training data folder has a series of images and training annotations.

Setup

• Install Tensorflow from https://www.tensorflow.org/install/
• Download the github repo for tensorflow models and place it in the top level folder
• Follow the Object Detection Installation instructions here

Run

Create a test.record file and a train.record file:

./generate_tfrecord --folder=path/to/data_dir --train_ratio=.70

Download COCO-pretrained Faster R-CNN with Resnet-101 model Unzip model in the data/models/ folder

Your data structure should look like

data/test.record data/train.record models/model/faster_rcnn_resnet101_coco_11_06_2017 models/model/train models/model/eval

In three seperate command line windows run:

export PYTHONPATH=$PYTHONPATH:pwd:pwd/slim

python models-master/research/object_detection/train.py     --logtostderr    
    --pipeline_config_path=./data/faster_rcnn_resnet101_cars.config  
    --train_dir=./data/models/model/train

python models-master/research/object_detection/eval.py     --logtostderr    
    --pipeline_config_path=./data/faster_rcnn_resnet101_cars.config    
    --checkpoint_dir=./data/models/model/train     
    --eval_dir=./data/models/model/eval

tensorboard --logdir=./data/models/model/

After a thousand or so steps you should be getting results. Look at your tensorflow eval images to gauge when to stop. FYI it took my 4770k Intel i7 about 24 hours to train.

Create a frozen version of your graph by selecting a checkpoint:

python object_detection/export_inference_graph.py \
    --input_type image_tensor \
    --pipeline_config_path /path/data/models/faster_rcnn_resnet101_cars.config \
    --trained_checkpoint_prefix /path/data/models/model/train/model.ckpt-18557 \
    --output_directory output_inference_graph.pb

Run saveimages.py to create a folder of images. This will save an image every second so leave it open as long as you want.

Analyze the images

analyzeimages \
        -path_images ./data/rawimages/ 
        -path_labels_map data/car_label_map.pbtxt 
        -save_directory data/processedimages/

Gotchas:

• When you start up an Amazon EC2 Instance using the AWS Deep Learning AMI you have to enable tensorflow