FashionAI Global Challenge: Key points Detection of Apparel

• Introduction: FashionAI Global Challenge
• Description: KeyPoints Detection of Apparel
• Leaderboard
• Score of this code : 4.25%
• Team ranked 32/2322
• Mainly base on Cascaded Pyramid Network for Multi-Person Pose Estimation.

Image data

Download

Folder Structure

• nets: store modified ResNet
• model: store checkpoint files
• outputs : store predicted files
• summary: store files for tensorboard
• train_set: place training data here
• test_set: place test data here
• fashion_evaluator.py: evaluator script
• fashion_generator.py: data generator
• fashion_helper.py: store a bunch of helper functions
• fashion_stacked.py: main file to define model
• train_script.py: train script
• test_script.py: test script

Prerequisites

Docker is recommended:

• nvidia-docker
• pull image from Docker Hub
  docker pull yd8534976/tf-aiden

Alternatively:

• python
• tensorflow-gpu (>= 1.4)
• numpy
• pandas
• opencv-python
• jupyter (optional)
• tensorboard (optional)

Pre-trained ResNet50 model

• You can download pre-trained models from tensorflow offical slim model zoo.
• Put checkpoint files into model/

Basic use

• Download datasets and put them into train_set/
• Download pre-trained ResNet-50 from slim model zoo
• Configure train_script.py
• Train end-to-end
  python train_script.py
• Visualize your training using TensorBoard
  tensorboard --logdir=summary/
• Generate predicted files
  python test_script.py
• Visualize your prediction using demo_notes.ipynb
• Evaluate your predictions
  python fashion_evaluator.py

Demo

• blouse
• skirt
• outwear
• dress
• trousers