The project was to detect traffic signs, based on the Megengine framework.

Overview

trafficsign

赛题

旷视AI智慧交通开源赛道,初赛1/177,复赛1/12。
本赛题为复杂场景的交通标志检测,对五种交通标志进行识别。

框架

megengine

算法方案

  • 网络框架

    • atss + resnext101_32x8d
  • 训练阶段

    • 图片尺寸
      最终提交版本输入图片尺寸为(1500,2100)

    • 多尺度训练(最终提交版本未采用)
      起初我们将短边设为(1024, 1056, 1088, 1120, 1152, 1184, 1216, 1248, 1280, 1312, 1344, 1376, 1408),随机选取短边后,长边按比例缩放,并使长边长度小于1800,从而进行多尺度训练,取得了很好的效果。 不过后期的mosaic和mixup在增强时对图片进行了缩放,实则隐含了多尺度训练,且效果优于上述方法,所以我们最终去掉了多尺度训练。

    • 数据增强

      • mosaic增强

        随机选择四张图片,对图片进行随机平移10%,尺度缩放(0.5,2.0),shear 0.1,最后将四张图片进行组合。

      • mixup增强

        随机选取两张图进行叠加,我们最终选用的比例是0.5 * 原图+0.5 * 新图片,同时其进行缩放(0.5,2.0)。

        下图为mosaic+mixup示例图:

        mosaic+mixup

      • 随机水平翻转

        直接对图片进行翻转,会导致第三个类别“arr_l”(左转线)和右转线混淆,故我们添加了class-aware的翻转,遇到有“arr_l”类的图片则不进行翻转。

      • 基于Albumentations库的各种增强(最终提交版本未采用)

        我们尝试了ShiftScaleRotate(验证集+0.5)、CLANE(验证集+1.0)、RandomBrightnessContrast等,但组合起来测试集提点欠佳,所以最后没用。

      • gridmask增强(最终提交版本未采用)

        生成一个和原图相同分辨率的mask(每个grid上全为0或全为1),然后将该mask与原图相乘得到一个图像。提点欠佳,所以没采用。

      • 类别平衡采样(最终提交版本未采用)

        使用类别平衡采样后,效果不是很好,这可能是因为数据集本身没有严重的类别不均衡。下面是我们统计的每个类别在图片中出现的频率。

        红灯 直行线 左转线 禁止行驶 禁止停车
        频率 0.356 0.228 0.201 0.257 0.485
  • 多尺度测试

    • 多尺度测试图片尺寸

      最后提交版本(2100,2700),(2100,2800),(2400,3200),如果继续增加尺度,map还会继续提高。

    • topk—nms

      对上述三个尺度生成的结果先进行nms,再将得到的结果框与剩下所有框进行topk—nms(保留与当前结果框iou大于0.85的topk的框,把这些框的坐标进行融合),参数设置vote_thresh=0.85, k=5。

  • 网络结构

    • 加上增强后,backbone从res50到res101再到resx101有稳定涨点。

    • 我们还在backbone部分尝试了dcn和gcnet,验证集收效甚微,最终没有采用。

模型训练与测试

  • 数据集位置
/path/to/ 
    |->traffic   
    |    |images     
    |    |annotations->|train.json     
    |    |             |val.json     
    |    |             |test.json      
  • 训练测试

在加上增强后,我们训练了36个epoch。

pip3 install --user -r requirements.txt

export PYTHONPATH=your_path/trafficsign:$PYTHONPATH

cd weights && wget https://data.megengine.org.cn/models/weights/atss_resx101_coco_2x_800size_45dot6_b3a91b36.pkl

python3 tools/train.py -n 4 -b 2 -f configs/atss_resx101_final.py -d your_datasetpath -w weights/atss_resx101_coco_2x_800size_45dot6_b3a91b36.pkl

python3 tools/test_final.py -n 4 -se 35 -f configs/atss_resx101_final.py -d your_datasetpath 

(-n 能抢到几张卡就写几吧qaq)

备注

以上提到的所有方法,无论最终是否采用,代码中均有实现。

感谢

https://github.com/MegEngine/Models/tree/master/official/vision/detection

https://github.com/MegEngine/YOLOX

Dynamic hair modeling from monocular videos using deep neural networks

Dynamic Hair Modeling The source code of the networks for our paper "Dynamic hair modeling from monocular videos using deep neural networks" (SIGGRAPH

53 Oct 18, 2022
Implementation of Kronecker Attention in Pytorch

Kronecker Attention Pytorch Implementation of Kronecker Attention in Pytorch. Results look less than stellar, but if someone found some context where

Phil Wang 16 May 06, 2022
Implementation of "Semi-supervised Domain Adaptive Structure Learning"

Semi-supervised Domain Adaptive Structure Learning - ASDA This repo contains the source code and dataset for our ASDA paper. Illustration of the propo

3 Dec 13, 2021
AQP is a modular pipeline built to enable the comparison and testing of different quality metric configurations.

Audio Quality Platform - AQP An Open Modular Python Platform for Objective Speech and Audio Quality Metrics AQP is a highly modular pipeline designed

Jack Geraghty 24 Oct 01, 2022
Generates all variables from your .tf files into a variables.tf file.

tfvg Generates all variables from your .tf files into a variables.tf file. It searches for every var.variable_name in your .tf files and generates a v

1 Dec 01, 2022
TensorFlow implementation of "Attention is all you need (Transformer)"

[TensorFlow 2] Attention is all you need (Transformer) TensorFlow implementation of "Attention is all you need (Transformer)" Dataset The MNIST datase

YeongHyeon Park 4 Jan 05, 2022
Multi-Horizon-Forecasting-for-Limit-Order-Books

Multi-Horizon-Forecasting-for-Limit-Order-Books This jupyter notebook is used to demonstrate our work, Multi-Horizon Forecasting for Limit Order Books

Zihao Zhang 116 Dec 23, 2022
学习 python3 以来写的一些垃圾玩具……

和东哥做兄弟 Author: chiupam 版权 未经本人同意,仓库内所有资源文件,禁止任何公众号、自媒体、开发者进行任何形式的转载、发布、搬运。 声明 这不是一个开源项目,只是把 GitHub 当作一个代码的存储空间,本项目不接受任何开源要求。 仅用于学习研究,禁止用于商业用途,不能保证其合法性

Chiupam 67 Mar 26, 2022
一个多模态内容理解算法框架,其中包含数据处理、预训练模型、常见模型以及模型加速等模块。

Overview 架构设计 插件介绍 安装使用 框架简介 方便使用,支持多模态,多任务的统一训练框架 能力列表: bert + 分类任务 自定义任务训练(插件注册) 框架设计 框架采用分层的思想组织模型训练流程。 DATA 层负责读取用户数据,根据 field 管理数据。 Parser 层负责转换原

Tencent 265 Dec 22, 2022
CenterNet:Objects as Points目标检测模型在Pytorch当中的实现

CenterNet:Objects as Points目标检测模型在Pytorch当中的实现

Bubbliiiing 267 Dec 29, 2022
Grow Function: Generate 3D Stacked Bifurcating Double Deep Cellular Automata based organisms which differentiate using a Genetic Algorithm...

Grow Function: A 3D Stacked Bifurcating Double Deep Cellular Automata which differentiates using a Genetic Algorithm... TLDR;High Def Trees that you can mint as NFTs on Solana

Nathaniel Gibson 4 Oct 08, 2022
Implementation of Shape and Electrostatic similarity metric in deepFMPO.

DeepFMPO v3D Code accompanying the paper "On the value of using 3D-shape and electrostatic similarities in deep generative methods". The paper can be

34 Nov 28, 2022
This repo provides the source code & data of our paper "GreaseLM: Graph REASoning Enhanced Language Models"

GreaseLM: Graph REASoning Enhanced Language Models This repo provides the source code & data of our paper "GreaseLM: Graph REASoning Enhanced Language

137 Jan 02, 2023
Music Classification: Beyond Supervised Learning, Towards Real-world Applications

Music Classification: Beyond Supervised Learning, Towards Real-world Applications

104 Dec 15, 2022
Residual Pathway Priors for Soft Equivariance Constraints

Residual Pathway Priors for Soft Equivariance Constraints This repo contains the implementation and the experiments for the paper Residual Pathway Pri

Marc Finzi 13 Oct 12, 2022
GNPy: Optical Route Planning and DWDM Network Optimization

GNPy is an open-source, community-developed library for building route planning and optimization tools in real-world mesh optical networks

Telecom Infra Project 140 Dec 19, 2022
PyTorch implementation of Masked Autoencoders Are Scalable Vision Learners for self-supervised ViT.

MAE for Self-supervised ViT Introduction This is an unofficial PyTorch implementation of Masked Autoencoders Are Scalable Vision Learners for self-sup

36 Oct 30, 2022
✔️ Visual, reactive testing library for Julia. Time machine included.

PlutoTest.jl (alpha release) Visual, reactive testing library for Julia A macro @test that you can use to verify your code's correctness. But instead

Pluto 68 Dec 20, 2022
Stacs-ci - A set of modules to enable integration of STACS with commonly used CI / CD systems

Static Token And Credential Scanner CI Integrations What is it? STACS is a YARA

STACS 18 Aug 04, 2022
[NeurIPS 2021] “Improving Contrastive Learning on Imbalanced Data via Open-World Sampling”,

Improving Contrastive Learning on Imbalanced Data via Open-World Sampling Introduction Contrastive learning approaches have achieved great success in

VITA 24 Dec 17, 2022