Object-Detection-In-Aerial-Images

Project Objectives

• To detect focal objects of interest in the satellite imagery
• To improve performance with architecture and framework
• Analysis on the result such as object size and type

Members

Name	Role	Task
차수연	Team member	EDA, Transform the dataset’s format, Training model and Analysis on the result and PPT. Mange GCP.
김하늘	Team leader	Manage the projects objectives. Training model and Analysis on the result.
유상민	Team member	Team member. Training model and Analysis on the result. Server management.
황동호	Team member	Team member. Training large Image and Analysis on the result. Issue management.

Dataset

• Dataset Download: AI hub
• Data Info
  • Images: png + tif
  • Annotation: json
  • Geographic context: Kml
  • Patch Size: 1,024x1,024
• Large Image
  • type: tif
  • Pixel Size: 12362 x 11344

Model Architecture

• RetinaNet

  

  • Feed Forward Network: ResNet
  • Backbone Network: FPN (Feature Pyramid Networks)
  • Loss function: Focal Loss

• Detectron2 Framework

  • Detectron2 is Facebook AI Research's next generation library that provides state-of-the-art detection and segmentation algorithms.
  • We trained a dectection model from an existing model pre-trained on the COCO dataset, available in detectron2's model zoo.
  • We chose R101 in RetinaNet baselines.

Problems and Solutions

Challenging tasks in Aerial Images

Object detection in aerial images is a challenging task due to the massive variations of scale, rotation, aspect ratio, and densely arranged targets.

More importantly, the lack of large-scale benchmarks has become a major obstacle to the development of object detection in aerial images.

The dataset contains 424,750 object instances of 21 categories of oriented-bounding-box annotations collected from 1,748 aerial images.

Based on this large-scale and well-annotated dataset from AI hub, we built baselines covering various state-of-the-art detection and segmentation algorithms with framework Detectron2, where the speed and accuracy performances of each model have been evaluated.

Problems and Solutions

1. Speed up model training

   • Chose the RetinaNet to match the speed of previous one-stage detectors while surpassing the accuracy of all existing state-of-the-art two-stage detectors

2. Address Class Imbalance

   • RetinaNet proposes the Focal Loss
     • Reshaping the standard cross entropy loss so that it down-weights the loss assigned to well-classified examples.
     • Focusing Training on a sparse set of hard examples and preventing the vast number of easy negatives from overwhelming the detector during training.

3. Data Imbalance

   • Separate the largest class from the others.
   • Train, evaluate and test two datasets separately.

4. Oriented Bounding Box

   

   The variations of the orientation of objects is caused by the bird's-eye view of aerial images. So we tried to predict rotated bounding box, but couldn’t succeed.

   Cause of failure

   • We transformed the Dota dataset format to Coco to use Detectron2 baselines. In this process, the produced annotations only included gemetric annotations and failed to produce angle annotations.
   • We focused on the speed and accuracy performances rather than predicting accurately the rotated bounding box. (Detectron2 provided various backbone networks, so we can customize datasets for our task.)

Results

• Metrics: Average Precision
  • The COCO Object Detection challenge also includes mean average recall as a detection metric. So we used average precision as a principal metric to evaluate object detectors. There are AP, AP50, AP75, mAP, AP@[0.5:0.95].

Dataset

Large Image

Training Step	AP	AP50	AP75
10,000	8.104	16.200	8.642
25,000	11.280	22.66	10.910
50,000	12.321	23.320	11.770
150,000	13.140	24.475	13.262

Paper Review

• 📃YOLO: You Only Look Once - YOLO v1~v3
• 📃YOLT: You Only Look Twice
• 📃RetinaNet

Link

Project Github

Presentation PPT