OffroadNet

OffroadNet is a collection of model architectures that focus on the task of improving the performance of traversable path detection on unpaved roads. We present the results of our experiments on the Robot Unstructured Ground Driving (RUGD) Dataset and train models to specifically focus on path detection performance. We use mIoU and mAcc as our metrics to compare the performance

of different models. OffroadNet (based on PSP- Net) and ResNet101 are one of the best combinations for offroad path detection because of it’s PPM module and fine-tuning based on the Cityscape dataset.