There exist good implementations of Faster R-CNN yet they lack support for recent ConvNet architectures. The aim of reproducing it from scratch is to fully utilize MXNet engines and parallelization for object detection.

[1] On Ubuntu 14.04.5 with device Titan X, cuDNN enabled. The experiment is VGG-16 end-to-end training.
[2] VGG network. Trained end-to-end on VOC07trainval+12trainval, tested on VOC07 test.
[3] VGG network. Fast R-CNN is the most memory expensive process.
[4] VGG network (parallelization limited by bandwidth). ResNet-101 speeds up from 2 img/s to 3.5 img/s.
[5] py-faster-rcnn does not support ResNet or recent caffe version.

• If you value stability and reproducibility over performance and efficiency, please refer to official implementations. There is no promise in all cases nor all experiments.
• If you value simplicity. Technical details are very complicated in MXNet. This is by design to attain maximum possible performance instead of patching fixes after fixes. Performance and parallelization are more than a change of parameter.
• If you want to do CPU training, be advised that it has not been verified yet. You will not encounter NOT_IMPLEMENTED_ERROR so it is still possible.
• If you are on Windows or Python3, some people reported it was possible with some modifications. But they have disappeared.

The above experiments were conducted at a mx-rcnn version using a MXNet fork, based on MXNet 0.9.1 nnvm pre-release.

• Prepare: bash script/additional_deps.sh
• Download training data: bash script/get_voc.sh
• Download pretrained model: bash script/get_pretrained_model.sh
• Training and testing: bash script/vgg_voc07.sh 0,1 (use gpu 0 and 1)

See if bash script/additional_deps.sh will do the following for you.

• Suppose HOME represents where this file is located. All commands, unless stated otherwise, should be started from HOME.
• Install python package cython easydict matplotlib scikit-image.
• Install MXNet version v0.9.5 or higher and MXNet Python Interface. Open python type import mxnet to confirm.
• Run make in HOME.

Command line arguments have the same meaning as in mxnet/example/image-classification.

• prefix refers to the first part of a saved model file name and epoch refers to a number in this file name. In model/vgg-0000.params, prefix is "model/vgg" and epoch is 0.
• begin_epoch means the start of your training process, which will apply to all saved checkpoints.
• Remember to turn off cudnn auto tune. export MXNET_CUDNN_AUTOTUNE_DEFAULT=0.

• An example of trained model (trained on VOC07 trainval) can be accessed from Baidu Yun (ixiw) or Dropbox. If you put the extracted model final-0000.params in HOME then use --prefix final --epoch 0 to access it.
• Try out detection result by running python demo.py --prefix final --epoch 0 --image myimage.jpg --gpu 0 --vis. Drop the --vis if you do not have a display or want to save as a new file.

The following tutorial is based on VOC data, VGG network. Supply --network resnet and --dataset coco to use other networks and datasets. Refer to script/vgg_voc07.sh and other experiments for examples.

See bash script/get_voc.sh and bash script/get_coco.sh will do the following for you.

• Make a folder data in HOME. data folder will be used to place the training data folder VOCdevkit and coco.
• Download and extract Pascal VOC data, place the VOCdevkit folder in HOME/data.
• Download and extract coco dataset, place all images to coco/images and annotation jsons to data/annotations.

(Skip this if not interested) All dataset have three attributes, image_set, root_path and dataset_path.

• image_set could be 2007_trainval or something like 2007trainval+2012trainval.
• root_path is usually data, where cache, selective_search_data, rpn_data will be stored.
• dataset_path could be something like data/VOCdevkit, where images, annotations and results can be put so that many copies of datasets can be linked to the same actual place.

See if bash script/get_pretrained_model.sh will do this for you. If not,

• Make a folder model in HOME. model folder will be used to place model checkpoints along the training process. It is recommended to set model as a symbolic link to somewhere else in hard disk.
• Download VGG16 pretrained model vgg16-0000.params from MXNet model gallery to model folder.
• Download ResNet pretrained model resnet-101-0000.params from ResNet to model folder.

See if bash script/vgg_alter_voc07.sh 0 (use gpu 0) will do the following for you.

• Start training by running python train_alternate.py. This will train the VGG network on the VOC07 trainval. More control of training process can be found in the argparse help.
• Start testing by running python test.py --prefix model/final --epoch 0 after completing the training process. This will test the VGG network on the VOC07 test with the model in HOME/model/final-0000.params. Adding a --vis will turn on visualization and -h will show help as in the training process.

See if bash script/vgg_voc07.sh 0 (use gpu 0) will do the following for you.

• Start training by running python train_end2end.py. This will train the VGG network on VOC07 trainval.#0000000000000000000000
• Start testing by running python test.py. This will test the VGG network on the VOC07 test.#0000000000000000000000

See if bash script/get_selective.sh and bash script/vgg_fast_rcnn.sh 0 (use gpu 0) will do the following for you.

• To reproduce Fast R-CNN, scipy is used to load selective search proposals.
• Download precomputed selective search data and place them to data folder. script/get_selective_search.sh will do this.
• Start training by running python -m rcnn.tools.train_rcnn --proposal selective_search to use the selective search proposal.
• Start testing by running python -m rcnn.tools.test_rcnn --proposal selective_search.
• script/vgg_fast_rcnn.sh will train Fast R-CNN on VOC07 and test on VOC07test.

Region Proposal Network solves object detection as a regression problem from the objectness perspective. Bounding boxes are predicted by applying learned bounding box deltas to base boxes, namely anchor boxes across different positions in feature maps. Training process directly learns a mapping from raw image intensities to bounding box transformation targets.

Fast R-CNN treats general object detection as a classification problem and bounding box prediction as a regression problem. Classifying cropped region feature maps and predicting bounding box displacements together yields detection results. Cropping feature maps instead of image input accelerates computation utilizing shared convolution maps. Bounding box displacements are simultaneously learned in the training process.

Faster R-CNN utilize an alternate optimization training process between RPN and Fast R-CNN. Fast R-CNN weights are used to initiate RPN for training. The approximate joint training scheme does not backpropagate rcnn training error to rpn training.

This repository provides Faster R-CNN as a package named rcnn.

• rcnn.core: core routines in Faster R-CNN training and testing.
• rcnn.cython: cython speedup from py-faster-rcnn.
• rcnn.dataset: dataset library. Base class is rcnn.dataset.imdb.IMDB.
• rcnn.io: prepare training data.
• rcnn.processing: data and label processing library.
• rcnn.pycocotools: python api from coco dataset.
• rcnn.symbol: symbol and operator.
• rcnn.tools: training and testing wrapper.
• rcnn.utils: utilities in training and testing, usually overloads mxnet functions.

This repository used code from MXNet, Fast R-CNN, Faster R-CNN, caffe, tornadomeet/mx-rcnn, MS COCO API.
Training data are from Pascal VOC, ImageNet, COCO.
Model comes from VGG16, ResNet.
Thanks to tornadomeet for end-to-end experiments and MXNet contributers for helpful discussions.

History of this implementation is:

• Fast R-CNN (v1)
• Faster R-CNN (v2)
• Faster R-CNN with module training (v3)
• Faster R-CNN with end-to-end training (v3.5, tornadomeet/mx-rcnn)
• Faster R-CNN with end-to-end training and module testing (v4)
• Faster R-CNN with accelerated training and resnet (v5)

mxnet/example/rcnn was v1, v2, v3.5 and now v5.

1. Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minjie Wang, Tianjun Xiao, Bing Xu, Chiyuan Zhang, and Zheng Zhang. MXNet: A Flexible and Efficient Machine Learning Library for Heterogeneous Distributed Systems. In Neural Information Processing Systems, Workshop on Machine Learning Systems, 2015
2. Ross Girshick. "Fast R-CNN." In Proceedings of the IEEE International Conference on Computer Vision, 2015.
3. Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. "Faster R-CNN: Towards real-time object detection with region proposal networks." In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016.
4. Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. "Caffe: Convolutional architecture for fast feature embedding." In Proceedings of the ACM International Conference on Multimedia, 2014.
5. Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. "The pascal visual object classes (voc) challenge." International journal of computer vision 88, no. 2 (2010): 303-338.
6. Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. "ImageNet: A large-scale hierarchical image database." In Computer Vision and Pattern Recognition, IEEE Conference on, 2009.
7. Karen Simonyan, and Andrew Zisserman. "Very deep convolutional networks for large-scale image recognition." arXiv preprint arXiv:1409.1556 (2014).
8. Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. "Deep Residual Learning for Image Recognition". In Computer Vision and Pattern Recognition, IEEE Conference on, 2016.
9. Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. "Microsoft COCO: Common Objects in Context" In European Conference on Computer Vision, pp. 740-755. Springer International Publishing, 2014.