Progressive Semi-Supervised Learning for Ground-to-Aerial Perception Knowledge Transfer

We introduce a novel progressive semi-supervised learning method for Ground-to-Aerial perception knowledge transfer with labeled images from the ground viewpoint and unlabeled images from flight viewpoints.

Results

Qualitative comparisons on the simulated data.

Qualitative comparisons on the real-world data.

Dependencies

• python 3.6.5
  
• torch 1.5.1
  
• torchvision 0.6.1
  
• tqdm 4.49.0
  
• matplotlib 3.3.2
  
• numpy 1.19.2
  
• pillow 8.4.0
  
• opencv-contrib-python 3.4.1.15
  

Datasets

We create two datasets for evaluation.

Structure of the datasets:

airs
├── uav0X (X denotes flying height)
     ├── images: directory storing images 
     ├── semantic: directory storing ground truth semantic maps
     ├── semantic_pl: directory storing pseudo-labeled semantic maps during training

├── pl_csv
     ├── airs.csv: a csv file storing the path of images and corresponding pseudo-labeled semantic maps during training

├── uav0X.csv (X denotes flying height): a csv file storing the path of images
├── uav01_labeled.csv: a csv file storing the path of images and semantic maps of ground viewpoint
├── uav0X_test_gt: a csv file storing the path of images and semantic maps of 2-9 meters for testing

airsim
├── car00: directory storing images and semantic maps of ground viewpoint

├── uav0X (X denotes flying height)
     ├── images: directory storing images 
     ├── semantic: directory storing ground truth semantic maps
     ├── semantic_pl: directory storing pseudo-labeled semantic maps during training

├── pl_csv
     ├── airsim.csv: a csv file storing the path of images and corresponding pseudo-labeled semantic maps during training

├── test (directory storing images and semantic maps of the test set)			
├── car00.csv: a csv file storing the path of images and semantic maps of ground viewpoint
├── uav0X.csv (X denotes flying height): a csv file storing the path of images and semantic maps
├── test.csv: a csv file storing the path of images and semantic maps of the test set

Running

The network of drone semantic segmentation is built on DeepLabv3+.
Download the datasets, unzip them to ./datasets/
Download our trained models, unzip them to ./runs/

• Test
  

  Testing on the AirSim-Drone: python test_airsim.py
  Testing on the AIRS-Street: python test_airs.py
  

• Train
  

  Training on the AirSim-Drone: python train_airsim.py
  Training on the AIRS-Street: python train_airs.py