TuSimple / mx-maskrcnn

An MXNet implementation of Mask R-CNN

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

MX Mask R-CNN

An MXNet implementation of Mask R-CNN.

This repository is based largely on the mx-rcnn implementation of Faster RCNN available here.



Main Results

Cityscapes

Method Training data Test data Average person rider car truck bus train motorcycle bicycle
Ours fine-only test 26.9 33.0 25.7 47.7 21.6 27.4 23.0 19.9 16.9
Reference[5] fine-only test 26.2 30.5 23.8 46.9 22.8 32.2 18.6 19.1 16.0
Ours fine-only val 31.3 32.6 26.6 49.5 26.5 45.4 32.1 17.6 20.4
Reference[5] fine-only val 31.5 - - - - - - - -
  • Backbone: Resnet-50-FPN

COCO

Coming soon, please stay tuned.

Requirement

We tested our code on:

Ubuntu 16.04, Python 2.7 with

numpy(1.12.1), cv2(2.4.9), PIL(4.3), matplotlib(2.1.0), cython(0.26.1), easydict

Preparation for Training

  1. Download Cityscapes data (gtFine_trainvaltest.zip, leftImg8bit_trainvaltest.zip). Extract them into 'data/cityscape/'. The folder structure would then look as shown below:
data/cityscape/
├── leftImg8bit/
│   ├── train/
│   ├── val/
│   └── test/
├── gtFine/
│   ├── train/
│   ├── val/
│   └── test/
└── imglists/
    ├── train.lst
    ├── val.lst
    └── test.lst
  1. Download Resnet-50 pretrained model.
bash scripts/download_res50.sh

  1. Build MXNet with ROIAlign operator.
cp rcnn/CXX_OP/* incubator-mxnet/src/operator/

To build MXNet from source, please refer to the tutorial.

  1. Build related cython code.
make
  1. Kick off training
bash scripts/train_alternate.sh

Preparation for Evaluation

  1. Prepare Cityscapes evaluation scripts.
bash scripts/download_cityscapescripts.sh
  1. Eval
bash scripts/eval.sh

Demo

  1. Download model, available at Dropbox/BaiduYun, and place it in the model folder.
  2. Make sure that you have the cityscapes data in 'data/cityscapes' folder.
bash scripts/demo.sh

Test single image

  1. Download model, available at Dropbox/BaiduYun, and place it in the model folder.
  2. Follow Preparation for Training (step1-step4)
  3. run bash scripts/demo_single_image.sh, you can change the image path in script demo_single_image.sh.

FAQ

Q: It says AttributeError: 'module' object has no attribute 'ROIAlign'.

A: This is because either

  • you forget to copy the operators to your MXNet folder

  • or you forget to re-compile MXNet and re-install MXNet python interface

  • or you install the wrong MXNet

    Please print mxnet.__path__ to make sure you use correct MXNet

Q: I encounter incubator-mxnet/mshadow/mshadow/././././cuda/tensor_gpu-inl.cuh:110: Check failed: err == cudaSuccess (7 vs. 0) Name: MapPlanKernel ErrStr:too many resources requested for launch at the begining.

A: Please try adding MSHADOW_CFLAGS += -DMSHADOW_OLD_CUDA=1 in mxnet/mshadow/make/mshadow.mk and re-compile MXNet.

References

  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. Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, Serge Belongie. "Feature Pyramid Networks for Object Detection." In Computer Vision and Pattern Recognition, IEEE Conference on, 2017.
  5. Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick. "Mask R-CNN." In Proceedings of the IEEE International Conference on Computer Vision, 2017.
  6. 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.
  7. 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.
  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. Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, Bernt Schiele. "The Cityscapes Dataset for Semantic Urban Scene Understanding." In Computer Vision and Pattern Recognition, IEEE Conference on, 2016.

About

An MXNet implementation of Mask R-CNN

License:Apache License 2.0


Languages

Language:Python 84.9%Language:C++ 7.5%Language:Cuda 4.4%Language:C 2.6%Language:Shell 0.6%Language:Makefile 0.1%