pedestrian-detector

Pedestrian Dectector using tensorflow Object detector API

Data Collection

139 of the data was taken from https://motchallenge.net/data/2D_MOT_2015/ , handpicked by me. I also scraped another 100 images from Google Images with the term pedestrian walking, from which I handpicked 71 images The images from Google introduces more noise to the training dataset, which can be both a good and bad thing. So we have in total 210 images in our training set

Note: One of the potential problem is that training data is skewed towards abled pedestrian, it might not be abled to detect people on wheelchairs etc

Data Annotation

A package named labelImg was used for ease of data annotation.

Note: to install this package, you must first build resource.py and resource.prc from source (following the instruction in accord with your OS) then copy them to python lib folder. Aftwards, labelImg can be easily installed like any other python packages pip install libImg

Getting it up and running

Installation

Install tf models following this doc Note: One of the dependency - Protobuf, must be exactly version 2.6.0 for it to work. Otherwise it will throw this error

Traceback (most recent call last):
  File "object_detection/train.py", line 49, in <module>
    from object_detection import trainer
  File "/Users/stanleynguyen/Documents/Projects/tf-models/research/object_detection/trainer.py", line 27, in <module>
    from object_detection.builders import preprocessor_builder
  File "/Users/stanleynguyen/Documents/Projects/tf-models/research/object_detection/builders/preprocessor_builder.py", line 21, in <module>
    from object_detection.protos import preprocessor_pb2
  File "/Users/stanleynguyen/Documents/Projects/tf-models/research/object_detection/protos/preprocessor_pb2.py", line 71, in <module>
    options=None, file=DESCRIPTOR),
TypeError: __init__() got an unexpected keyword argument 'file'

Training

Command to start training locally (mainly for testing unless we have very powerful GPUs)

python3 object_detection/train.py \
--logtostderr \
 --pipeline_config_path=../../pedestrian-detector/training/ssd_inception_v2_coco.local.config \
 --train_dir=../../pedestrian-detector/training/

Command to deploy it yo Google ML Engine Note:

• Before deployment, there are some ML Engine problems that need to be addressed, please see this comment
• The instance type of ML engine should be set to large_model to afford our training

# copy all training data to gs
gsutil cp -r training gs://pedestrian-detector/training
gsutil cp -r data gs://pedestrian-detector/data
# creating the  required packages
# you must be inside models/research
python3 setup.py sdist && cd slim && python3 setup.py sdist
# copy all config to gs
gsutil cp -r training gs://pedestrian-detector/training
# copy all training data to gs
gsutil cp -r data gs://pedestrian-detector/data
# deploy!
# you are still inside models/research
# running training job
gcloud ml-engine jobs submit training pedestrian_detection_`date +%s` --runtime-version 1.2 --job-dir=gs://pedestrian-detector --packages dist/object_detection-0.1.tar.gz,slim/dist/slim-0.1.tar.gz --module-name object_detection.train --region us-central1 --config ../../pedestrian-detector/training/cloud.yml -- --train_dir=gs://pedestrian-detector --pipeline_config_path=gs://pedestrian-detector/training/ssd_inception_v2_coco.cloud.config
# running eval job
gcloud ml-engine jobs submit training pedestrian_detection_`date +%s` --runtime-version 1.2 --job-dir=gs://pedestrian-detector --packages dist/object_detection-0.1.tar.gz,slim/dist/slim-0.1.tar.gz --module-name object_detection.eval --region us-central1 --scale-tier BASIC_GPU -- --checkpoint_dir=gs://pedestrian-detector/train --eval_dir=gs://pedestrian-detector/eval --pipeline_config_path=gs://pedestrian-detector/training/ssd_inception_v2_coco.cloud.config
  505  history

Monitoring

# train
tensorboard --logdir=gs://pedestrian-detector/train --debug
# eval
tensorboard --logdir=gs://pedestrian-detector/eval --debug