1202kbs/object-detection-api-wrapper

Simple wrapper for Google Object Detection API

Google Object Detection API Wrapper

Simple wrapper functions for Google Object Detection API.

Instructions

Method 1

1. Download the repository
2. Create a 'detection_model_zoo' directory in the extracted repository file
3. Download model from Tensorflow detection model zoo
4. Copy the downloaded model into 'detection_model_zoo' directory
5. Create a file containing the urls of images to be analyzed. This wrapper supports both .csv and .json extensions. (Click link for example format)
6. Check the Jupyter notebook for further instructions and examples

Method 2

1. Download the repository
2. Select the model to use from Tensorflow detection model zoo
3. Create a file containing the urls of images to be analyzed. This wrapper supports both .csv and .json extensions. (Click link for example format)
4. Run detect_cm.py in command line with --downloaded False. Check below section for more detail

Running in Command Line

python detect_cm.py --model-name MODEL_NAME --url-file URL_FILE --extension EXTENSION --downloaded DOWNLOADED --do-rescale DO_RESCALE

Example:

python detect_cm.py --model-name faster_rcnn_inception_resnet_v2_atrous_coco_11_06_2017 --url-file Imagenet_sample_images --extension .json --downloaded True --do-rescale False

Required Arguments

• --model-name : Name of the model to use. No Default
• --url-file : Name of the url file without .json extension. No Default
• --extension : Url file file type, .json or .csv. Default: .csv
• --downloaded : Indicate whether you have already downloaded model .tar file or not. Wrapper will automatically create 'detection_model_zoo' directory and download model file if set to False. Default: False
• --do-rescale : Indicate whether you want to rescale the images or not. Default: False

Optional Arguments

• --width-threshold : If rescale is set to True, all the images with width size over width threshold will be rescaled to fit width threshold. Default: 500
• --json-output-file : Name of the json output file to dump results. Default: output.json
• --n-threads : Number of threads to use. Default: 64
• --visualize : Indicate whether you want to visualize the results. Default: False
• --labels : Path to the label file. Default: object_detection/data/mscoco_label_map.pbtxt

Run python detect_cm.py --help to see a list of all options.

Successful execution will create a .json file which contains the results of object detection. Contents of the .json file can be read into a dictionary of numpy arrarys with json2dict function in helpers.py. For additional functions for visualization or modification of the results of object detection, look into helpers.py.

Running within IDE

1. Import detect.py
2. Run detect.object_detect(...)
3. The object_detect() function takes the same arguments as command line execution. Look into detect.py for more details

If the execution is successful, the function will return a dictionary of numpy arrays containing the object detection results and create a .json file which contains the results of object detection. Contents of the .json file can be read into a dictionary of numpy arrarys with json2dict function in helpers.py. For additional functions for visualization or modification of the results of object detection, look into helpers.py.

Example:

import detect

model_name = 'faster_rcnn_inception_resnet_v2_atrous_coco_11_06_2017'
url_file = 'Imagenet_sample_images'
extension = '.csv'
downloaded = True
do_rescale = True
width_threshold = 1000
visualize = True

res = detect.object_detect(model_name=model_name, url_file=url_file, extension=extension, downloaded=downloaded, do_rescale=do_rescale, width_threshold=width_thresholdm visualize=visualize)

Etc.

• Check the function docstring in helpers.py for more info
• .proto files have already been compiled with Protobuf compiler (Protoc)
• This code has been tested with Tensorflow 1.1 on Windows 10

Helpful Links

• Original link for Google Tensorflow Object Detection API
• Protobuf compiler download link

Prerequisites

• Python 3.5.x
• Tensorflow-gpu >= 1.1
• Scikit-image
• Matplotlib
• Pandas
• Numpy

Updates

• 2017/07/04 : Added image rescale function. For some reason, rescaling drops the recognition accuracy by a significant amount, so using this is not recommended until resolved.