Tensorflow implementation of RPN(Region Proposal Networks) for single class object detection. The setting of this project refer to this paper so that it can be used for pedestrian detection(other tasks need to change parametres).

Installtion

1. clone this repository
2. install Tensorflow

Datasets

The current version only supports Pascal VOC format. It means you must provide images and corresponding labels(.xml).

Your directory structure should be similar to :

|-- DATASET_DIR
    |-- annotations
    |-- images

Then you can use tf_convert_data.py script to convert data to TF-Records:

python tf_convert_data.py \
    --dataset_dir=/path/to/your/DATASET_DIR \
    --output_dir=/paht/to/your/OUTPUT_DIR

Both training data and test data can be get in this way.

Training

The script train_ssd_network.py is in charged of training the network. This project supports fine-tune with pre-trained VGG-16 model. You can download it from here. Then you can begin your train with train_rpn.py script:

python train_rpn.py \
    --train_dir=/path/where/to/write/training/log \
    --dataset_dir=/path/to/DATASET_DIR \
    --model_path=/path/to/vgg_16.ckpt

It model_path is set to None, the project will not use any pre-trained model and begin to train from random initialization.