English | 简体中文

SegNet

SegNet

• SegNet
  • 一、Introduction
  • 二、Accuracy
  • 三、Dataset
  • 四、Environment
  • 五、Quick Start
    • Quick Train：
    • Quick Infer：
  • 六、Code structure and More explaination
  • 七、Model information

一、Introduction

SegNet，a model for semantic segmentation
Paper：https://github.com/yassouali/pytorch-segmentation
Opensourced on AI Studio , can work online here：https://aistudio.baidu.com/aistudio/projectdetail/2293857

二、Accuracy

On the dataset camvid 11classses, miou = 0.601

三、Dataset

Dataset used: camvid
Download link：https://aistudio.baidu.com/aistudio/datasetdetail/79232
This version come from：https://www.kaggle.com/naureenmohammad/camvid-dataset, in which resolution is low (480x360)，id of labels range from 0 to 11

Dataset Size：

• train_num：367
• val_num：101
• test_num：233

Format as follows：

• test test_imgs（.png）
• testannot test_labels（.png）
• train train_imgs（.png）
• trainannot train_labels（.png）
• val val_imgs（.png）
• valannot val_labels（.png）

四、Environment

• Hardware
  • GPU
  • CPU
• Framework / Software
  • PaddlePaddle = 2.1
  • PaddleSeg

五、Quick Start

Quick Train：

cd PaddleSeg
python train.py \
       --config config.yml \
       --do_eval \
       --use_vdl \
       --save_interval 500 \
       --save_dir output

Parameters explaination：

• config ：root of config.yml
• do_eval ：validate while saving model
• use_vdl ：use VisualDl (a tool like tensorboard)
• save_interval ：frequency to save model
• save_dir ：directory to save models and logs

Quick Infer：

python data/PaddleSeg/predict.py \
       --config config.yml \
       --model_path output_bs_8——pre/best_model/model.pdparams \
       --image_path data/PaddleSeg/camvid/test \
       --save_dir output/result

Parameters explaination：

• config ：root of config.yml
• model_path ：root of model to use
• image_path ：path of imgs needed to infer
• save_dir ：path to save outputs

六、Code structure and More explaination

This project is powered by PaddlePaddle ,code structrure is like PaddleSeg https://github.com/PaddlePaddle/PaddleSeg，to use, we just need to set the config.yml file.

Introduction to config.yml ：

batch_size: 4  #batch_size
iters: 1000    #iterations

train_dataset: #train_dataset config
  type: Dataset #type of train_dataset
  dataset_root: data/PaddleSeg/camvid #list of train_dataset
  train_path: data/PaddleSeg/camvid/train_list.txt #list of train_dataset
  num_classes: 12 #classed（background counts as well）
  transforms: #data pre-execution and enhancement
    - type: Resize #resize before put into network
      target_size: [512, 512] #resize to 512*512
    - type: RandomHorizontalFlip # use RandomHorizontalFlip to enhance data
    - type: Normalize #Normalize data
  mode: train

val_dataset: #train_dataset config
  type: Dataset #type of train_dataset
  dataset_root: data/PaddleSeg/camvid #list of train_dataset
  val_path: data/PaddleSeg/camvid/val_list.txt #list of train_dataset
  num_classes: 12 #classed（background counts as well）
  transforms: #data pre-execution and enhancement
    - type: Resize  #resize before put into network
      target_size: [512, 512]  #resize to 512*512
    - type: Normalize #Normalize img
  mode: val

optimizer: #set optimizer
  type: sgd #adopt SGD（Stochastic Gradient Descent）as optimizer
  momentum: 0.9 #set momentum
  weight_decay: 4.0e-5 #weight_decay, to avoid overfitting

learning_rate: #set learning_rate
  value: 0.1  #initial learning_rate
  decay:
    type: poly  #adopt poly as means of m
    power: 0.9  #rate of decay
    end_lr: 0   #end learning_rate

loss: #set loss funtion
  types:
    - type: CrossEntropyLoss #loss function type
  coef: [1]

model: #model config 
  type: SegNet  #set model category
  num_classes: 12

七、Model information

information	description
Author	chen jiejun
E-mail	stuartchen2018@outlook.com
Date	2021.8.8
Framework version	PaddlePaddle = 2.1
Application scenarios	Semantic Segmantation
Support hardware	GPU CPU