Exploring Data Variance challenges in Fusion of Radar and Camera for Robotics and Autonomous Driving

摘要

Inductive bias is a fervently targeted area in the recent developments of deep learning. However, the variance problems are often unattended and are proceeding with the launch of a new dataset. The variance and addressing its problems to achieve better results for a particular task is of importance which follows from the advocacy of the no free lunch theorem. The hard inductive bias of the hierarchical feature pyramid network is not sufficient to generalize and rely on the necessary information in images to produce robust object detection. This work reduced the impact of the inducive variance presented in camera and radar fusion for object detection in harsh environmental conditions. We explore the variance in radar data with the inductive bias of convolution pooling. The proposed architecture uses the untransformed radar data and extracts features using convolution pooling. Experimentations are performed on the nuScenes dataset which shows a 5% mAP improvement in results compared to the baseline. It is observed that the improvement in results is based on choices of data, hyper-parameters, and network composition. The qualitative results of object detection show the significance of proposed tweaks for robotics and autonomous driving applications.