CVPR 2019

Yihui He, Chenchen Zhu, Jianren Wang, Marios Savvides, Xiangyu Zhang, Carnegie Mellon University & Megvii Inc.

Large-scale object detection datasets (e.g., MS-COCO) try to define the ground truth bounding boxes as clear as possible. However, we observe that ambiguities are still introduced when labeling the bounding boxes. In this paper, we propose a novel bounding box regression loss for learning bounding box transformation and localization variance together. Our loss greatly improves the localization accuracies of various architectures with nearly no additional computation. The learned localization variance allows us to merge neighboring bounding boxes during non-maximum suppression (NMS), which further improves the localization performance. On MS-COCO, we boost the Average Precision (AP) of VGG-16 Faster R-CNN from 23.6% to 29.1%. More importantly, for ResNet-50-FPN Mask R-CNN, our method improves the AP and AP90 by 1.8% and 6.2% respectively, which significantly outperforms previous state-of-the-art bounding box refinement methods.

If you find the code useful in your research, please consider citing:

@inproceedings{klloss,
  title={Bounding Box Regression with Uncertainty for Accurate Object Detection},
  author={He, Yihui and Zhu, Chenchen and Wang, Jianren and Savvides, Marios and Zhang, Xiangyu },
  booktitle={2019 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2019},
  organization={IEEE}
}

Please find installation instructions for Caffe2 and Detectron in INSTALL.md.

When installing cocoapi, please use my fork to get AP80 and AP90 scores.

Inference without Var Voting (8 GPUs):

python2 tools/test_net.py \
    --cfg configs/e2e_faster_rcnn_R-50-FPN_2x.yaml

You will get:

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.385
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.578
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.412
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.209
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.412
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.515
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.323
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.499
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.522
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.321
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.553
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.680
 Average Precision  (AP) @[ IoU=0.60      | area=   all | maxDets=100 ] = 0.533
 Average Precision  (AP) @[ IoU=0.70      | area=   all | maxDets=100 ] = 0.461
 Average Precision  (AP) @[ IoU=0.80      | area=   all | maxDets=100 ] = 0.350
 Average Precision  (AP) @[ IoU=0.85      | area=   all | maxDets=100 ] = 0.269
 Average Precision  (AP) @[ IoU=0.90      | area=   all | maxDets=100 ] = 0.154
 Average Precision  (AP) @[ IoU=0.95      | area=   all | maxDets=100 ] = 0.032

Inference with Var Voting:

python2 tools/test_net.py \
    --cfg configs/e2e_faster_rcnn_R-50-FPN_2x.yaml \
    STD_NMS True

You will get:

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.392
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.576
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.425
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.212
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.417
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.526
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.324
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.528
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.564
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.346
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.594
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.736
 Average Precision  (AP) @[ IoU=0.60      | area=   all | maxDets=100 ] = 0.536
 Average Precision  (AP) @[ IoU=0.70      | area=   all | maxDets=100 ] = 0.472
 Average Precision  (AP) @[ IoU=0.80      | area=   all | maxDets=100 ] = 0.363
 Average Precision  (AP) @[ IoU=0.85      | area=   all | maxDets=100 ] = 0.281
 Average Precision  (AP) @[ IoU=0.90      | area=   all | maxDets=100 ] = 0.165
 Average Precision  (AP) @[ IoU=0.95      | area=   all | maxDets=100 ] = 0.037

python2 tools/train_net.py \
    --cfg configs/e2e_faster_rcnn_R-50-FPN_2x.yaml

Please create a new issue.

Detectron is Facebook AI Research's software system that implements state-of-the-art object detection algorithms, including Mask R-CNN. It is written in Python and powered by the Caffe2 deep learning framework.

At FAIR, Detectron has enabled numerous research projects, including: Feature Pyramid Networks for Object Detection, Mask R-CNN, Detecting and Recognizing Human-Object Interactions, Focal Loss for Dense Object Detection, Non-local Neural Networks, Learning to Segment Every Thing, Data Distillation: Towards Omni-Supervised Learning, DensePose: Dense Human Pose Estimation In The Wild, and Group Normalization.

Example Mask R-CNN output.

The goal of Detectron is to provide a high-quality, high-performance codebase for object detection research. It is designed to be flexible in order to support rapid implementation and evaluation of novel research. Detectron includes implementations of the following object detection algorithms:

• Mask R-CNN -- Marr Prize at ICCV 2017
• RetinaNet -- Best Student Paper Award at ICCV 2017
• Faster R-CNN
• RPN
• Fast R-CNN
• R-FCN

using the following backbone network architectures:

• ResNeXt{50,101,152}
• ResNet{50,101,152}
• Feature Pyramid Networks (with ResNet/ResNeXt)
• VGG16

Additional backbone architectures may be easily implemented. For more details about these models, please see References below.

• 4/2018: Support Group Normalization - see GN/README.md

Detectron is released under the Apache 2.0 license. See the NOTICE file for additional details.

If you use Detectron in your research or wish to refer to the baseline results published in the Model Zoo, please use the following BibTeX entry.

@misc{Detectron2018,
  author =       {Ross Girshick and Ilija Radosavovic and Georgia Gkioxari and
                  Piotr Doll\'{a}r and Kaiming He},
  title =        {Detectron},
  howpublished = {\url{https://github.com/facebookresearch/detectron}},
  year =         {2018}
}

We provide a large set of baseline results and trained models available for download in the Detectron Model Zoo.

Please find installation instructions for Caffe2 and Detectron in INSTALL.md.

After installation, please see GETTING_STARTED.md for brief tutorials covering inference and training with Detectron.

To start, please check the troubleshooting section of our installation instructions as well as our FAQ. If you couldn't find help there, try searching our GitHub issues. We intend the issues page to be a forum in which the community collectively troubleshoots problems.

If bugs are found, we appreciate pull requests (including adding Q&A's to FAQ.md and improving our installation instructions and troubleshooting documents). Please see CONTRIBUTING.md for more information about contributing to Detectron.

• Data Distillation: Towards Omni-Supervised Learning. Ilija Radosavovic, Piotr Dollár, Ross Girshick, Georgia Gkioxari, and Kaiming He. Tech report, arXiv, Dec. 2017.
• Learning to Segment Every Thing. Ronghang Hu, Piotr Dollár, Kaiming He, Trevor Darrell, and Ross Girshick. Tech report, arXiv, Nov. 2017.
• Non-Local Neural Networks. Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Tech report, arXiv, Nov. 2017.
• Mask R-CNN. Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. IEEE International Conference on Computer Vision (ICCV), 2017.
• Focal Loss for Dense Object Detection. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. IEEE International Conference on Computer Vision (ICCV), 2017.
• Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour. Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Tech report, arXiv, June 2017.
• Detecting and Recognizing Human-Object Interactions. Georgia Gkioxari, Ross Girshick, Piotr Dollár, and Kaiming He. Tech report, arXiv, Apr. 2017.
• Feature Pyramid Networks for Object Detection. Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017.
• Aggregated Residual Transformations for Deep Neural Networks. Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017.
• R-FCN: Object Detection via Region-based Fully Convolutional Networks. Jifeng Dai, Yi Li, Kaiming He, and Jian Sun. Conference on Neural Information Processing Systems (NIPS), 2016.
• Deep Residual Learning for Image Recognition. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.
• Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Conference on Neural Information Processing Systems (NIPS), 2015.
• Fast R-CNN. Ross Girshick. IEEE International Conference on Computer Vision (ICCV), 2015.