Accelerating Deep Neural Networks for Efficient Scene Understanding in Multi-Modal Automotive Applications

Environment perception constitutes one of the most critical operations performed by semi-and fully- autonomous vehicles. In recent years, Deep Neural Networks (DNNs) have become standard tool for solutions owing to their impressive capabilities in analyzing and modelling complex dynamic scenes, from (often muti-modal) sensory inputs. However, well-established performance DNNs comes at cost increased time storage complexity, which may problematic automotive systems due requirement a short prediction horizon (as many cases inference must be real time) limited computational, storage, energy resources mobile systems. A common way addressing this problem is transform original large pre-trained networks into new smaller models, utilizing Model Compression Acceleration (MCA) techniques, improving both execution efficiency. Within MCA framework, paper, we investigate application two state-of-the-art weight-sharing namely Vector Quantization (VQ) Dictionary Learning (DL) one, as well novel extensions, towards acceleration compression widely used 2D 3D object-detection applications. Apart individual (uni-modal) networks, also present evaluate multi-modal late-fusion algorithm combining detection results detectors. Our evaluation studies are carried out on KITTI Dataset. The obtained lend themselves twofold interpretation. On hand, they showcase significant gains that can achieved via weight sharing selected DNN detectors, with accuracy loss, highlight differences between utilized approaches. other, demonstrate substantial boost outcome unimodal using proposed based approach. Indeed, our experiments shown, pairing high-performance DL-based technique loss-mitigating effect fusion approach, leads highly accelerated models (up approximately 2.5× 6× respectively) loss fused ranging within single-digits figures low around 1% class “cars”).