Using Dynamic Pruning Technique for Efficient Depth Estimation for Autonomous Vehicles

Abstract

Even with the significant progress that has been achieved in monocular depth estimation in recent years, the need for better real-time inference and reduction in computing resources usage associated with the network performance is persistent. In this paper, an enquiry into the efficacy of pruning on depth estimation models is performed. Encoder-decoder model based on the ResNet-50 backbone architecture employing pruning based on channel prioritization is designed to achieve higher performance and prediction speed. This is while attempting to keep a balance in the trade-off between accuracy and performance of the network. The presented approach is trained and evaluated for outdoor scenery on the KITTI dataset to demonstrate the effectiveness and the performance improvement of the presented framework when compared to similar methods. This shows competitive accuracy when compared to state-of-the-art methods and highlights how pruning can speed up inference time by more than 16% and leading to fewer operations compared to the non-pruned model.