Domain adaptive Drone 2 maps - GeoAI
- 0 Collaborators
Digitizing an accurate map can be a very cumbersome tasks which humans have been doing since all long. There has been advances in automatically digitizing the maps, but they are not that accurate. The problem is that the imagery accumulated from satellite are low resolution and smaller features are tough to distinguish, hence creating a inaccurate map. Not only, the satellite imagery is low resolution but is also very expensive to gather and label. The problem of spatial inconsistency also makes the digitizing maps of different location a challenge. In this project we try to solve these issues using a Domain Adaptive Image Segmentation using high resolution imagery provided by Intel Drones. ...learn more
Project status: Concept
Robotics, RealSense™, Artificial Intelligence
Groups
Student Developers for AI
Intel Technologies
AI DevCloud / Xeon,
MKL,
Intel Opt ML/DL Framework,
Intel Python,
Movidius NCS
Overview / Usage
- The drone will be flown over an unknown spatial area(given an extent) where it will capture high resolution imagery, infrared imagery, and other information captured from sensors like LiDAR.
- All of this information will be combined and segmented by a pre-trained model for high resolution image segmentation.
- The domain will be adapted to the domain of the images the model is trained on then the image will be segmented.
- All the raster data will be converted back to geo-saptial data and will be updated on a GIS platform.
Methodology / Approach
- Image Segmentation will be done using U-Net. It is a Convolutional Neural Network based Image segmentation model. It has feature pyramid network for high-res feature transfer from earlier layers to later layers.
- Domain adaptation will done using CycleGANs with a modified loss function which will help in Image Segmentation.
Technologies Used
- PyTorch
- FastAI
- Intel Python
- Intel mkl
- Intel Falcon 8+
- Intel AI DevCloud
- Intel NCS