Skip to content

Dataset for training lane segmentation and traffic sign detection from Goodgame, the Champion of FPT Digital Race - Cuộc Đua Số 2020.

License

Notifications You must be signed in to change notification settings

makerviet/via-dataset-cds-2020

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

18 Commits
 
 
 
 
 
 
 
 

Repository files navigation

VIA Dataset

The repository contains:

  • Procedures to use VIA Dataset with two tasks: Object detection and semantic segmentation.
  • How to implement into your custom dataset.

Download datasets

Documentation

Object detection

Object detection folder contains: dataset folder, two files .csv (test.csv, train.csv). The train.csv has 12,764 images and the test.csv has 2561 images.

Object Detection 
    │─── Data
         |───000000_10.png
         |───000001_10.png
         |─── ...
    │─── test.csv
    │─── train.csv

The structure of .csv:

filename xmin ymin xmax ymax class_id
00072.jpg 148 53 159 63 4

In object detection task, we have 6 traffic signs (6 classes): Turn left, Turn Right, Straight, Stop, No Turn Left, No Turn Right.

Semantic segmentation

Setup sementation data folders

Segmentation
    │─── GGDataSet
         |─── train_frames
             |─── train
                 |─── train_000001.png
         |─── train_masks
             |─── train
                 |─── train_000001.png         
         |─── val_frames
             |─── val
                 |─── val_000001.png         
         |─── val_masks
             |─── val
                 |─── val_000001.png         
         |─── label_colors.txt
    │─── model_pb
    │─── models
    │─── train.py
    │─── convert_pb.py

VIA segmentation dataset has 6,240 training images, 1,448 validation images. In our task, we have to predict 3 classes: Background, Line, Road. The label_colors.txt contains RGB color code of classes and we handle it to convert classes into one-hot vector.

How to use VIA Dataset

Object detection

In VIA experiment, we implement Single Shot Detection (SSD). You do not need to follow our instructions if you want to handle the data for only your purposes.

  1. Upload VIA Dataset to Google Drive
  2. Upload object_detection.ipynb to Google Colab
  3. Modify your dataset locate path in Google Drive and your dataset path link to images and .csv files.

You can run notebook in local with requirements: Keras version 2.2.4, Tensorflow version 1.15 and git clone this repository:

git clone https://github.com/pierluigiferrari/ssd_keras

Semantic segmentation

In VIA experiment, we implement PSPNet and use combine-loss is Dice Loss and Focal Loss. We use a library segmentation, you can read the docs to modify segmentation architecture.

To train segmentation task by VIA dataset, run the following commands

# Keras 2.2.4, Tensorflow >= 1.15

pip install -U segmentation-models

# Modify dataset path in train.py

img_dir = 'your_path/GGDataSet/'

DATA_PATH = 'your_path/GGDataSet/'

# Train

python3 train.py

# If you want to run pretrained model faster, you need to convert model to frozen graph 

python3 convert_pb.py

How to implement into your custom dataset

In object detection task, we use the labelimg tool and the labelme tool to label segmentation dataset.

To object detection, we need to convert .xml files to .csv as (train.csv above). This xml_to_csv.py to help you handle it, remember that in our .csv we only contains (filename, xmin,ymin,xmax,ymax,class_id).

To semantic segmentation, the labelme tool export .json, we need to convert .json files to .png. Run the following commands

git clone https://github.com/wkentaro/labelme.git

cd labelme/examples/semantic_segmentation

./labelme2voc.py data_annotated data_dataset_voc --labels labels.txt

You can see the label PNG file in data_dataset_voc/SegmentationClassPNG/ folder. Modify data_annotated folder to your_dataset.

Website: https://via.makerviet.org/

About

Dataset for training lane segmentation and traffic sign detection from Goodgame, the Champion of FPT Digital Race - Cuộc Đua Số 2020.

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Contributors 3

  •  
  •  
  •