NEXTLab Car Detection Project

Example

Main Contributers

• Jeongtae Shin(@Klassikcat): Project Management, Model build, Engineering
• Jaewoong Lee(@colin19950703): Augmentation, model build, Enigeering
• Hongkyun Kim(@ghdrbs0302): Data Inspection, Test data Labeling, Engineering

Installation

!pip install -U git+https://github.com/Klassikcat/project-NEXTLab-CNN-EfficientNet

pre-trained Weights

https://drive.google.com/file/d/1UyVnR5Pi7NvzD3LPnkllGEyn8afMi5jF/view?usp=sharing

Structure

project-NEXTLab
┖ core
  ┖ config.py               # Tensorflow Configuration
  ┖ utils.py
┖ model
  ┖ EfficientNet.py         # Model block and builder
  ┖ initializers.py         # from EfficientNetv2 (https://arxiv.org/abs/2104.00298)
  ┖ layer.py                # from EfficietnNetv2
  ┖ model.py                # from EfficientNetv2
  ┖ params.py               # from EfficientNetv2
┖ utils
  ┖ augmentation            # Data augmentation tools
    ┖ augmentation.py
    ┖ launcher_augment.py
    ┖ launcher_classname.py
  ┖ from_folder             
    ┖ loader.py             # get train/valid/test dataset from directory
  ┖ tfrecord                 
    ┖ tfrecordMaker         # Make TFRecord
    ┖ tfrecordViewer        # Visualize, Load TFRecord 
  ┖ unicodeUtil.py          # Korean utilities assemble jamo
  ┖ visuialization.py       # Plots(data distribution, heatmap)
setup.py                    # installer
readme.md

Stack

Objective

• Making EfficientNetv2 application that distinguish 322 cars by using photos captured by traffic camera
• Traffic enforcement, Parking lot management.

Data

Train/Validatation Data

• Uses modified "자동차 차종/연식/번호판 인식용 영상" from AIHub: https://aihub.or.kr/aidata/27727
• Uses 522 classes in "세단", "SUB", "해치백", "승합"

Test Data

• Road Camera dataset from NEXTLab: https://www.nextlab.ai/
• 5016 images belonging to 80 classes

Data Distribution

approximately 82% of data is korean cars(hyundai, KIA)

approximately 75% of data is korean cars(sonata, etc...)

51% sedan, 28.8% SUV, 10% Hatchback, 9% van

mainly consists with 2010~2018 cars

Data Augmentation

• use these augmentations taking the possible road situations into account.
• Drop the class with less than 30 pieces of data in order to avoid data leakage
• randomly pick 300 data. if the class has less than 30 data, proceed Data Augmentation for equal distribution of data

Model Structure

• EfficientNetB0(224, 224) with 30 layer freeze

Train_validation accuracy/loss

top 1 accuracy/loss

• train: 0.9051/0.2755
• valid: 0.8960/0.3769

top 3 accuracy

• train: 0.9880
• valid: 0.9717

Further Work

[ ] Increase valid/test accuracy by making Ensemble model

[ ] Label Smoothing

[ ] Fix all false label