AI-cup-2024-spring

Image Data Generation Competition

Follow the steps to reproduce our results.

• Team ID: TEAM_5574
• Place: 21(Public), 14 (Private)
• Member:
  • 藍翊庭, NCKU (Leader)
  • 周韋恩, NCKU
  • 呂軒名, NCKU
  • 連思涵, NCKU

Introduction

Our task is to translate the black-and-white draft image into drone imagery.

Draft Image	Drone Image

Dataset Preprocess

• Remove Low-Quality Images:

  We first analyze the dataset and saw some low-quality images. We then remove the low-quality images. some low-quality images are:

  Low Quality Draft Image	Low Quality Drone Image

• Split Dataset into 2 domains: (ROAD and RIVER)

  We extract the dataset from the raw training dataset and split it into two domains: ROAD and RIVER.

  dataset
  ├── train_ROAD
  │   ├── trainA (Draft Images)
  │   └── trainB (Drone Images)
  └── train_RIVER
      ├── trainA (Draft Images)
      └── trainB (Drone Images)
  

Model Pipeline

Hyperparameters

Training

We train the 2 model with the following hyperparameters:

• Epochs: 1000 (n_epochs=900, n_epochs_decay=100)
• netG = unet_256
• Batch Size: 1

Testing

After training, We get the net_G for each model and use it to generate the drone image from the draft image. We eventually choose 550_net_G.pth for ROAD model and 550_net_G.pth for RIVER model.

Evaluation Metric

• FID (Fréchet Inception Distance) : 計算真實影像和生成影像之特徵距離，越低表示圖像品質越好。
• Final Score : 河流影像與道路影像會個別計算一個 FID 分數，FIDriver, FIDroad 分別加權評分後得到最終分數。

Results

Use the Final score to evaluate the performance about each models.

Model	Public Testing	Private Testing
Baseline	249.76	247.94
Enhanced	132.95	131.11

Setup

To reach our results, you can follow the steps below:

1. run dataset/preprocess_dataset.ipynb
2. run train_model.ipynb (optional, we have provided the pretrained model)
3. run test_model.ipynb