Skip to content

Latest commit

 

History

History
73 lines (55 loc) · 5.97 KB

README.md

File metadata and controls

73 lines (55 loc) · 5.97 KB

DetectoRS: Detecting Objects with Recursive Feature Pyramid and Switchable Atrous Convolution

Abstract

Many modern object detectors demonstrate outstanding performances by using the mechanism of looking and thinking twice. In this paper, we explore this mechanism in the backbone design for object detection. At the macro level, we propose Recursive Feature Pyramid, which incorporates extra feedback connections from Feature Pyramid Networks into the bottom-up backbone layers. At the micro level, we propose Switchable Atrous Convolution, which convolves the features with different atrous rates and gathers the results using switch functions. Combining them results in DetectoRS, which significantly improves the performances of object detection. On COCO test-dev, DetectoRS achieves state-of-the-art 55.7% box AP for object detection, 48.5% mask AP for instance segmentation, and 50.0% PQ for panoptic segmentation.

Citation

We provide the config files for DetectoRS: Detecting Objects with Recursive Feature Pyramid and Switchable Atrous Convolution.

@article{qiao2020detectors,
  title={DetectoRS: Detecting Objects with Recursive Feature Pyramid and Switchable Atrous Convolution},
  author={Qiao, Siyuan and Chen, Liang-Chieh and Yuille, Alan},
  journal={arXiv preprint arXiv:2006.02334},
  year={2020}
}

Dataset

DetectoRS requires COCO and COCO-stuff dataset for training. You need to download and extract it in the COCO dataset path. The directory should be like this.

mmdetection
├── mmdet
├── tools
├── configs
├── data
│   ├── coco
│   │   ├── annotations
│   │   ├── train2017
│   │   ├── val2017
│   │   ├── test2017
|   |   ├── stuffthingmaps

Results and Models

DetectoRS includes two major components:

  • Recursive Feature Pyramid (RFP).
  • Switchable Atrous Convolution (SAC).

They can be used independently. Combining them together results in DetectoRS. The results on COCO 2017 val are shown in the below table.

Method Detector Lr schd Mem (GB) Inf time (fps) box AP mask AP Config Download
RFP Cascade + ResNet-50 1x 7.5 - 44.8 config model | log
SAC Cascade + ResNet-50 1x 5.6 - 45.0 config model | log
DetectoRS Cascade + ResNet-50 1x 9.9 - 47.4 config model | log
RFP HTC + ResNet-50 1x 11.2 - 46.6 40.9 config model | log
SAC HTC + ResNet-50 1x 9.3 - 46.4 40.9 config model | log
DetectoRS HTC + ResNet-50 1x 13.6 - 49.1 42.6 config model | log
DetectoRS HTC + ResNet-101 20e 19.6 50.5 43.9 config model | log

Note: This is a re-implementation based on MMDetection-V2. The original implementation is based on MMDetection-V1.