Implements Faster R-CNN Architecture
https://medium.com/@venkysai.96/faster-r-cnn-object-detection-7e48e5b9a906
pip install requirements.txt
To train the model, please run
python src/train.py
Once visom is installed, to run it , open a terminal and type visdom
The model can be visualized at http://localhost:8097
The dataset can be downloaded from the following links :-
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar
The dataset can be extracted and stored in the parent directory. If not, its location can be changed in src/utils/config.py
at voc_data_dir
The directory structure is as follows :-
- data : contains the necessary files needed for loading the VOC dataset along with transformation functions.
- dataset : base class which instantiates the voc dataset and transforms the raw data.
- util : helper functions for preprocessing the image, bounding boxes
- voc_dataset : contains class needed to process/parse the voc dataset data
- models : this contains the faster rcnn models and all of its constituent methods.
- faster_rcnn : core model with train, predict functions. Instantiates and Calls all other models (head, rpn).
- head : contains the methods needed for VGG head, ROI pooling of faster rcnn.
- rpn : contains the methods needed for calling region proposal network.
- utils : this contains the methods needed for faster rcnn models.
- anchors : this has all the utility functions related to anchors.
- config : contains the configuration/options.
- helper : helper methods
- proposals : used to generate rpn layer and its corresponding ground truth proposals.
- visualization : visualization utility functions
https://github.com/chenyuntc/simple-faster-rcnn-pytorch
You can contribute in serveral ways such as creating new features, improving documentation etc.
MIT Licence