shineyruan / faster-rcnn-toy

Simple implementation of Faster-RCNN

faster-rcnn-toy

Simple implementation of Faster-RCNN

CIS 680: Advanced Topics in Machine Perception FA2020, University of Pennsylvania.

Authors: Zhihao Ruan [e-mail], Kun Huang [e-mail]

Introduction

Faster-RCNN is a deep neural network targeting real-time object detection. It consists of a backbone made up with an FPN (Feature Pyramid Network) and an RPN (Region Proposal Network). The outputs coming from RPN are fed into a Box Head, which is an object detection head consisting of a classification branch and a regression branch, and each of them are set up with a set of fully connected layers. This repo implements a simplified version of Faster-RCNN, which uses a pre-trained FPN and RPN as backbone, and implements the Box Head. For infos about Faster-RCNN, please refer to the official code.

Object Detection Demos

On a simplified COCO dataset, we reached an mAP of 0.5265.

Usage

Please download all the 3 folders from the Google Drive link and put them all under the project root directory.

As a course project, we used a simplified COCO dataset provided from Prof. Jianbo Shi of University of Pennsylvania, which as attached in data/ and dataset building methods can be found in src/dataset.py. All of the Box Head architecture are defined in src/BoxHead.py. To run the training pipeline, please run the following commands:

cd ${PROJECT_ROOT}
python src/main_train.py

To run the validation pipeline, please run

cd ${PROJECT_ROOT}
python src/main_infer.py

Configurations

We used Anaconda virtual environment as the official development environment. All the Conda packages used can be found in environment.yml.

References

@misc{ren2016faster,
      title={Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks},
      author={Shaoqing Ren and Kaiming He and Ross Girshick and Jian Sun},
      year={2016},
      eprint={1506.01497},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}