Geospatial object detection and segmentation
Dataset: NWPU-VHR-10
Model: Mask R-CNN with torchvision
mAP (IOU = 0.5): 0.973
mAR (max detections per image = 100): 0.770
This application is deployed with two replicated sets on a Kubernetes cluster
Things need to do next:
- Evaluate FPS (Frames per seconds)
- Try to use a light weight weight model (e.g., mobilenet_v2, suqeezeNet…) to replace ResNet to improve the FPS
- Compare current metric with the work done by others
- Use a confusion matrix to recognize patterns which classes are mixed with which or which classes cannot be detect
- May need more samples for some classes to improve the recall.