Augmenting Anchors by the Detector Itself

Introduction

It is difficult to determine the scale and aspect ratio of anchors for anchor-based object detection methods. Current state-of-the-art object detectors either determine anchor parameters according to objects' shape and scale in a dataset, or avoid this problem by utilizing anchor-free method. In this paper, we propose a gradient-free anchor augmentation method named AADI, which means Augmenting Anchors by the Detector Itself. AADI is not an anchor-free method, but it converts the scale and aspect ratio of anchors from a continuous space to a discrete space, which greatly alleviates the problem of anchors' designation. Furthermore, AADI does not add any parameters or hyper-parameters, which is beneficial for future research and downstream tasks. Extensive experiments on COCO dataset show that AADI has obvious advantages for both two-stage and single-stage methods, specifically, AADI achieves at least 2.1 AP improvements on Faster R-CNN and 1.6 AP improvements on RetinaNet, using ResNet-50 model. We hope that this simple and cost-efficient method can be widely used in object detection.

• For RPN

  • Baseline

    Num anchors	AR100	AR1000	ARs	ARm	ARl
    1	45.5	55.6	31.4	52.8	60.0
    3	45.7	58.0	31.4	52.7	61.1

  • Ablation Study

    dilation	Anchor Guided	AR100	AR1000	ARs	ARm	ARl
    1		52.8	60.6	40.2	60.8	63.6
    2		54.8	64.7	39.0	63.1	70.6
    2	√	56.3	66.7	39.5	64.9	73.4
    3		53.7	64.0	35.4	62.1	73.9
    3	√	55.6	67.6	36.1	64.3	77.6
    4		52.2	60.5	30.9	61.3	76.6
    4	√	54.4	65.5	33.0	63.7	78.9

• For RetinaNet

  • Ablation Study

    AADI	dilation	AP	AP50	AP75	APs	APm	APl
    	1	38.2	58.4	41.1	24.3	42.2	48.5
    √	1	37.3	56.4	40.2	22.0	39.9	46.8
    √	2	39.8	57.5	43.5	22.1	43.5	50.6
    √	3	38.3	54.6	41.7	20.0	43.1	51.1

  • With IoU

    AP	AP50	AP75	APs	APm	APl
    40.2	57.7	43.8	24.1	43.1	52.2

  • With 3x schedule (RetinaNet with giou, AADI with smooth l1)

    Model	AP	AP50	AP75	APs	APm	APl
    RetinaNet	39.6	59.3	42.2	24.9	43.3	50.7
    AADI-RetinaNet	41.4	59.3	45.2	24.8	44.9	54.0

• For Faster R-CNN

  • Ablation Study

    AADI	dilation	AP	AP50	AP75	APs	APm	APl	FPS
    	1(3 anchors)	37.9	58.8	41.1	22.4	41.1	49.1	26.3
    √	2	40.3	59.3	44.3	24.2	43.3	52.2	22.4
    √	3	40.8	59.5	45.0	24.0	44.6	53.1	22.4
    √	4	40.5	58.7	44.6	23.2	44.8	52.7	22.3

  • 3x schedule

    Backbone	AP	AP50	AP75	APs	APm	APl	FPS
    R-50 FPN	42.5	61.2	46.5	25.3	46.2	55.5	22.6
    DCN-50 FPN	44.1	63.1	48.2	28.3	46.9	58.4	20.1
    R-101 FPN	44.5	63.2	48.7	26.9	48.3	57.4	17.4

• Detectron2

Detectron2 is Facebook AI Research's next generation library that provides state-of-the-art detection and segmentation algorithms. It is the successor of Detectron and maskrcnn-benchmark. It supports a number of computer vision research projects and production applications in Facebook.

Installation

See installation instructions.

Getting Started

See Getting Started with Detectron2, and the Colab Notebook to learn about basic usage.

Learn more at our documentation.

Citing Detectron2

@misc{wu2019detectron2,
  author =       {Yuxin Wu and Alexander Kirillov and Francisco Massa and
                  Wan-Yen Lo and Ross Girshick},
  title =        {Detectron2},
  howpublished = {\url{https://github.com/facebookresearch/detectron2}},
  year =         {2019}
}

@misc{wan2021augmenting,
      title={Augmenting Anchors by the Detector Itself}, 
      author={Xiaopei Wan and Shengjie Chen and Yujiu Yang and Zhenhua Guo and Fangbo Tao},
      year={2021},
      eprint={2105.14086},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}