competition-oiltank-detection

Competition for oil tank detection in satellite images.

• Host : Korea Aerospace Research Institute (KARI)
• Organizer : AI Factory
• Link : https://aifactory.space/task/2454/overview
• Team name : 3M1B

Problem definition

Data

• 2D Color Image (1024x1024)
• Satellite Image

Task

• Object detection

  • Single object : only one class (oil tank)
  • Oriented object detection : bounding box is rotated

• Metric

  • Average Precision (AP)
  • Intersection over Union (IoU) > 0.5

Challenge

• Bounding Box is rotated (different from COCO)
• Very small data size (Only 70 Images)
  • Pretrained weight is needed & Hard to validate
  • Fine tuning is difficult
• Oil tanks vary in size

Approach

Model

• Yolov5

  • Good
    • It is not SOTA but verified and stable → User friendly and nice docs
    • Easy to apply for custom data → Time saving
  • Bad
    • No simple option for using oriented object detection
    • Less models than MMdetection

• Bigger model

  • Usually the bigger the model, the higher the performance
  • Does it worked?
    • Yolov5x (the biggest) outperformed yolov5m (the middle among models)
    • Bigger model forces to use of smaller batch sizes due to the memory.
      However, the advantage of a bigger model seems to exceed the disadvantage of smaller batch size

• Optimal balance between image size & batch size

  • Since the model has batch normalization, bigger batch might be helpful
  • Due to the GPU memory restriction, the bigger the batch, the smaller the image size
  • Originally, yolov5 was trained by 640x640 size image but competition gave us 1024 image
  • Does it worked?
    • Increase image size and sacrifices batch size was better
    • Reducing image size to 640 showed low performance.
      Reducing image size might be made it harder to detect small size oil tanks

• Augmentation and training technique

  • Multiscale learning : using += 50% image size while training
    • It forces to use smaller batch size due to the increased size while training (16 → 8)
  • Cosine annealing
    • Not sure it worked because we couldn't see
  • Does it worked?
    • No, performance was decreased
    • There was no clear difference of validation score bewteen using this technique or not.
      We suspect generalization performance was worsened
    • Reduced batch size due to the memory is presumed to be a reason of bad generalization performance

Data

• Ignore bounding box rotation
  • Why?
    • Yolo pretrained weight is based on COCO (No rotation)
    • Oil tank bounding box closes to the square shape due to its nature
  • How ?
    • Center of mass of polygon + Square shape bounding box assumption
    • Width and height are equal and calcuated by square root of bounding box area
  • Does it worked?
    • I worked but performance is not high
    • It is well matched to the square-shaped nature of oil tank (In terms of IoU)

What we learned

• How to use Yolo(v5) for custom data
  • A non-COCO format data to COCO format
  • Still, for DOTA dataset, MMdetection could be a better choice
• How to improve performance
  • Using bigger model
  • Consider carefully when using increasing image size technique due to the trade off