The surprising impact of mask-head architecture on novel class segmentation

被引:10
作者
Birodkar, Vighnesh [1 ]
Lu, Zhichao [1 ]
Li, Siyang [1 ]
Rathod, Vivek [1 ]
Huang, Jonathan [1 ]
机构
[1] Google, Mountain View, CA 94043 USA
来源
2021 IEEE/CVF INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV 2021) | 2021年
关键词
D O I
10.1109/ICCV48922.2021.00693
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Instance segmentation models today are very accurate when trained on large annotated datasets, but collecting mask annotations at scale is prohibitively expensive. We address the partially supervised instance segmentation problem in which one can train on (significantly cheaper) bounding boxes for all categories but use masks only for a subset of categories. In this work, we focus on a popular family of models which apply differentiable cropping to a feature map and predict a mask based on the resulting crop. Under this family, we study Mask R-CNN and discover that instead of its default strategy of training the mask-head with a combination of proposals and groundtruth boxes, training the mask-head with only groundtruth boxes dramatically improves its performance on novel classes. This training strategy also allows us to take advantage of alternative mask-head architectures, which we exploit by replacing the typical mask-head of 2-4 layers with significantly deeper off-the-shelf architectures (e.g. ResNet, Hourglass models). While many of these architectures perform similarly when trained in fully supervised mode, our main finding is that they can generalize to novel classes in dramatically different ways. We call this ability of mask-heads to generalize to unseen classes the strong mask generalization effect and show that without any specialty modules or losses, we can achieve state-of-the-art results in the partially supervised COCO instance segmentation benchmark. Finally, we demonstrate that our effect is general, holding across underlying detection methodologies (including anchor-based, anchor-free or no detector at all) and across different backbone networks. Code and pre-trained models are available at https://git.io/deepmac.
引用
收藏
页码:6995 / 7005
页数:11
相关论文
共 56 条
  • [1] [Anonymous], 2019, P IEEE CVF C COMP VI
  • [2] Large-scale interactive object segmentation with human annotators
    Benenson, Rodrigo
    Popov, Stefan
    Ferrari, Vittorio
    [J]. 2019 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2019), 2019, : 11692 - 11701
  • [3] Bochkovskiy A., 2020, PREPRINT, DOI DOI 10.48550/ARXIV.2004.10934
  • [4] Cascade R-CNN: High Quality Object Detection and Instance Segmentation
    Cai, Zhaowei
    Vasconcelos, Nuno
    [J]. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 2021, 43 (05) : 1483 - 1498
  • [5] Carion N., 2020, ARXIV200512872
  • [6] Chandrashekar M, 2018, IEEE INT CONF BIG DA, P4906, DOI 10.1109/BigData.2018.8622438
  • [7] CenterNet: Keypoint Triplets for Object Detection
    Duan, Kaiwen
    Bai, Song
    Xie, Lingxi
    Qi, Honggang
    Huang, Qingming
    Tian, Qi
    [J]. 2019 IEEE/CVF INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV 2019), 2019, : 6568 - 6577
  • [8] Enze Xie, 2020, 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Proceedings, P12190, DOI 10.1109/CVPR42600.2020.01221
  • [9] Scalable Object Detection using Deep Neural Networks
    Erhan, Dumitru
    Szegedy, Christian
    Toshev, Alexander
    Anguelov, Dragomir
    [J]. 2014 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR), 2014, : 2155 - 2162
  • [10] The PASCAL Visual Object Classes Challenge: A Retrospective
    Everingham, Mark
    Eslami, S. M. Ali
    Van Gool, Luc
    Williams, Christopher K. I.
    Winn, John
    Zisserman, Andrew
    [J]. INTERNATIONAL JOURNAL OF COMPUTER VISION, 2015, 111 (01) : 98 - 136