Augmenting Anchors by the Detector Itself

Related tags

Computer Visionaadi
Overview

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}
}
Repository for Scene Text Detection with Supervised Pyramid Context Network with tensorflow.

Scene-Text-Detection-with-SPCNET Unofficial repository for [Scene Text Detection with Supervised Pyramid Context Network][https://arxiv.org/abs/1811.0

121 Oct 15, 2021
2 telegram-bots: for image recognition and for text generation

💻 📱 Telegram_Bots 🔎 & 📖 2 telegram-bots: for image recognition and for text generation. About Image recognition bot: User sends a photo and bot de

Marina Polukoshko 1 Jan 27, 2022
Markup for note taking

Subtext: markup for note-taking Subtext is a text-based, block-oriented hypertext format. It is designed with note-taking in mind. It has a simple, pe

Gordon Brander 224 Jan 01, 2023
Basic functions manipulating images using the OpenCV library

OpenCV Basic functions manipulating images using the OpenCV library. Reading Ima

Shatha Siala 3 Feb 17, 2022
Tracking the latest progress in Scene Text Detection and Recognition: Must-read papers well organized

SceneTextPapers Tracking the latest progress in Scene Text Detection and Recognition: must-read papers well organized Information about this repositor

Shangbang Long 763 Jan 01, 2023
A version of nrsc5-gui that merges the interface developed by cmnybo with the architecture developed by zefie in order to start a new baseline that is not heavily dependent upon Python processing.

NRSC5-DUI is a graphical interface for nrsc5. It makes it easy to play your favorite FM HD radio stations using an RTL-SDR dongle. It will also displa

61 Dec 22, 2022
An Agnostic Computer Vision Framework - Pluggable to any Training Library: Fastai, Pytorch-Lightning with more to come

An Agnostic Object Detection Framework IceVision is the first agnostic computer vision framework to offer a curated collection with hundreds of high-q

airctic 790 Jan 05, 2023
Image Recognition Model Generator

Takes a user-inputted query and generates a machine learning image recognition model that determines if an inputted image is or isn't their query

Christopher Oka 1 Jan 13, 2022
WACV 2022 Paper - Is An Image Worth Five Sentences? A New Look into Semantics for Image-Text Matching

Is An Image Worth Five Sentences? A New Look into Semantics for Image-Text Matching Code based on our WACV 2022 Accepted Paper: https://arxiv.org/pdf/

Andres 13 Dec 17, 2022
Read-only mirror of https://gitlab.gnome.org/GNOME/ocrfeeder

================================= OCRFeeder - A Complete OCR Suite ================================= OCRFeeder is a complete Optical Character Recogn

GNOME Github Mirror 81 Dec 23, 2022
Pixie - A full-featured 2D graphics library for Python

Pixie - A full-featured 2D graphics library for Python Pixie is a 2D graphics library similar to Cairo and Skia. pip install pixie-python Features: Ty

treeform 65 Dec 30, 2022
Maze generator and solver with python

Procedural-Maze-Generator-Algorithms Check out my youtube channel : Auctux Ressources Thanks to Jamis Buck Book : Mazes for programmers Requirements P

Joseph 19 Dec 07, 2022
Repository of conference publications and source code for first-/ second-authored papers published at NeurIPS, ICML, and ICLR.

Repository of conference publications and source code for first-/ second-authored papers published at NeurIPS, ICML, and ICLR.

Daniel Jarrett 26 Jun 17, 2021
This is a pytorch re-implementation of EAST: An Efficient and Accurate Scene Text Detector.

EAST: An Efficient and Accurate Scene Text Detector Description: This version will be updated soon, please pay attention to this work. The motivation

Dejia Song 544 Dec 20, 2022
PyTorch Re-Implementation of EAST: An Efficient and Accurate Scene Text Detector

Description This is a PyTorch Re-Implementation of EAST: An Efficient and Accurate Scene Text Detector. Only RBOX part is implemented. Using dice loss

365 Dec 20, 2022
End-to-end pipeline for real-time scene text detection and recognition.

Real-time-Scene-Text-Detection-and-Recognition-System End-to-end pipeline for real-time scene text detection and recognition. The detection model use

Fangneng Zhan 89 Aug 04, 2022
Natural language detection

Detect the language of text. What’s so cool about franc? franc can support more languages(†) than any other library franc is packaged with support for

Titus 3.8k Jan 02, 2023
huoyijie 1.2k Dec 29, 2022
ocroseg - This is a deep learning model for page layout analysis / segmentation.

ocroseg This is a deep learning model for page layout analysis / segmentation. There are many different ways in which you can train and run it, but by

NVIDIA Research Projects 71 Dec 06, 2022