Second place winning implementation in Speed-Hack: Autonomous Driving @ Daimler. Object detection, instance segmentation and tracking on Cityscapes.
For object detection and instance segmentation, we use a pre-trained Mask R-CNN network with a ResNet50 backbone. Based on the detected bounding boxes, a simple object tracking was implemented using the intersection over union.
To install dependencies use:
pip install -r requirements.txt