Chainer_Mask_R-CNN
Chainer implementation of Mask R-CNN - the multi-task network for object detection, object classification, and instance segmentation.
(https://arxiv.org/abs/1703.06870)
ζ₯ζ¬θͺη README
What's New
- Training result for R-50-C4 model has been evaluated!
- COCO box AP = 0.346 using our trainer (0.355 with official boxes)
- COCO mask AP = 0.287 using our trainer (0.314 with official boxes)
Examples
- to be updated
Requirements
- Chainer
- Chainercv
- Cupy
(operable if your environment can run chainer > v3 with cuda and cudnn.)
(verified as operable: chainer==3.1.0, chainercv==0.7.0, cupy==1.0.3)
$ pip install chainer
$ pip install chainercv
$ pip install cupy
- Python 3.0+
- NumPy
- Matplotlib
- OpenCV
TODOs
- Precision Evaluator (bbox, COCO metric)
- Detectron Model Parser
- Modify ROIAlign
- Mask inference using refined ROIs
- Precision Evaluator (mask, COCO metric)
- Improve segmentation AP for R-50-C4 model
- Feature Pyramid Network (R-50-FPN)
- Keypoint Detection (R-50-FPN, Keypoints)
Benchmark Results
| Box AP 50:95 | Segm AP 50:95 | |
| Ours (1 GPU) | 0.346 | 0.287 |
| Detectron model | 0.350 | 0.295 |
| Detectron caffe2 | 0.355 | 0.314 |
Inference with Pretrained Models
- Download the pretrained model from the [Model Zoo] (https://github.com/facebookresearch/Detectron/blob/master/MODEL_ZOO.md)
(modellink ofR-50-C4 MaskatEnd-to-End Faster & Mask R-CNN Baselines) - Make
modelfilesdirectory and put the downloaded filemodel_final.pklin it - Execute:
python utils/detectron_parser.py
- And the converted model file is saved in
modelfiles - Run the demo:
python demo.py --bn2affine --modelfile modelfiles/e2e_mask_rcnn_R-50-C4_1x_d2c.npz --image <input image>
Prerequisites for training
-
Download 'ResNet-50-model.caffemodel' from the "OneDrive download" of ResNet pretrained models for model initialization and place it in ~/.chainer/dataset/pfnet/chainer/models/
-
COCO 2017 dataset : the COCO dataset can be downloaded and unzipped by:
bash getcoco.sh
Setup the COCO API:
git clone https://github.com/waleedka/coco
cd coco/PythonAPI/
make
python setup.py install
cd ../../
note: the official coco repository is not python3 compatible.
Use the repository above in order to run our evaluation.
Train
python train.py
arguments and the default conditions are defined as follows:
'--dataset', choices=('coco2017'), default='coco2017'
'--extractor', choices=('resnet50','resnet101'), default='resnet50', help='extractor network'
'--gpu', '-g', type=int, default=0
'--lr', '-l', type=float, default=1e-4
'--batchsize', '-b', type=int, default=8
'--freeze_bn', action='store_true', default=False, help='freeze batchnorm gamma/beta'
'--bn2affine', action='store_true', default=False, help='batchnorm to affine'
'--out', '-o', default='result', help='output directory'
'--seed', '-s', type=int, default=0
'--roialign', action='store_true', default=True, help='True: ROIAlign, False: ROIpooling'
'--step_size', '-ss', type=int, default=400000
'--lr_step', '-ls', type=int, default=480000
'--lr_initialchange', '-li', type=int, default=800
'--pretrained', '-p', type=str, default='imagenet'
'--snapshot', type=int, default=4000
'--validation', type=int, default=30000
'--resume', type=str
'--iteration', '-i', type=int, default=800000
'--roi_size', '-r', type=int, default=14, help='ROI size for mask head input'
'--gamma', type=float, default=1, help='mask loss balancing factor'
note that we use a subdivision-based updater to enable training with large batch size.
Demo
Segment the objects in the input image by executing:
python demo.py --image <input image> --modelfile result/snapshot_model.npz --contour
Evaluation
Evaluate the trained model with COCO metric (bounding box, segmentation) :
python train.py --lr 0 --iteration 1 --validation 1 --resume <trained_model>
Citation
Please cite the original paper in your publications if it helps your research:
@article{DBLP:journals/corr/HeGDG17,
author = {Kaiming He and
Georgia Gkioxari and
Piotr Doll{\'{a}}r and
Ross B. Girshick},
title = {Mask {R-CNN}},
journal = {CoRR},
volume = {abs/1703.06870},
year = {2017},
url = {http://arxiv.org/abs/1703.06870},
archivePrefix = {arXiv},
eprint = {1703.06870},
timestamp = {Wed, 07 Jun 2017 14:42:32 +0200},
biburl = {http://dblp.org/rec/bib/journals/corr/HeGDG17},
bibsource = {dblp computer science bibliography, http://dblp.org}
}