endernewton/tf-faster-rcnn

This repository is a TensorFlow implementation of Faster RCNN, a well-known approach to object detection, which means teaching a computer to identify and locate specific objects (like cars, people, or animals) within photographs. The code was written by a researcher at Carnegie Mellon University and is based on an earlier Python version that used a different deep learning framework called Caffe. The README opens with a notice that this project is now deprecated and points readers toward a more current implementation in a separate project called TensorPack. Object detection systems like Faster RCNN work by first proposing regions in an image that might contain objects, then classifying what those regions contain and refining the boundaries. This repository supports three families of neural network architectures for doing that work: VGG16, ResNet (in sizes 50, 101, and 152), and MobileNet. The code was tested against standard benchmark datasets used in computer vision research, namely the Pascal VOC dataset and the COCO dataset, and the README reports the detection accuracy scores achieved on those benchmarks. Beyond the core detection logic, the implementation adds a few conveniences for researchers: the ability to run validation checks during training to watch for signs the model is overfitting, the ability to stop and resume a training run from a saved checkpoint, and automatic logging of loss statistics and network behavior to TensorBoard, a visualization tool that displays training progress as charts in a browser. Setting the project up requires a working TensorFlow installation at version r1.2, several Python packages including OpenCV for image processing, and a GPU for any practical training run. The README walks through compiling some required low-level code, downloading pre-trained model weights, and running the training and testing scripts. This repository is aimed at machine learning researchers and students who want to experiment with or build on the Faster RCNN architecture using TensorFlow.