Here is a list of ideas for getting started and improving baselines on the Airborne Object Detection challenge -
Using YOLO models for object detection and tracking
- Out of the box YOLOv3 model with DeepSort algorithm for tracking
- YOLO model fine tuned on the Airprime dataset
Detectron2 provides numerous models for object detection and segmentation and a flexible library for adding new ones. This can be a good starting point for building detection models.
Using JDE based algorithms for object tracking
Most algorithms explored are in SDE (Separate detection and embedding) paradigm. Recent algorithms in the Joint Detection and Embedding (JDE) paradigm have achieved superior performance in MOT leaderboards. Few notable examples -
- FAIR MOT
- GSDT - JDE+GNN based detection, using graphs to connect objects across temporal and spatial dimensions.
PaddlePaddle provides a framework with a few MultiObject tracking methods (deepsort, fairmot, jde), and a flexible interface for adding new ones.
A few suggestions for improving object detection
- Downscaling the input resolution to the model might make it hard to detect smaller airborne objects. Scaling up your input resolution to the model could help.
- Using a model with high resolution would make your model large and increase inference time. Alternatively, you could tile your images as a preprocessing step and continue using smaller models. (The Power of Tiling for Small Object Detection)
- Removing birds from predictions can help avoid extra false positives, since we are not interested in alerting birds.