Non-residual unrestricted pruned ultra-faster line detection for edge devices

Abstract

Line detection with deep learning is a popular visual task that focuses mostly on lane detection. It requires quicker inference speed and lower consumption, especially for high-speed edge device applications. Based on the UFAST, we propose the Non-Residual Unrestricted Pruned Ultra-faster (NRUPU) line detection via a novel model compression method including non-interference structural reconstruction (NISR), shallow channel priority reservation (SCPR) pruning and non-residual equivalent transformation (NRET). NISR is a structure reconstruction scheme allocating residual branches into each layer to solve the cross-layer channel interference in ResNet-18. SCPR pruning directly uses the factors of BN layers to build channel importance evaluation for backbone and designs channel selection method for head based on data distribution consistency, reducing the parameters of each layer independently. Then NRET losslessly converts the multi-branch model to a single-branch one containing only convolution, linear, and relu, which reduces implementation complexity on edge devices. These designs follow the theoretical foundations: the data distribution transformation trend and effect of gradient back-propagation on model learning ability. Compared with previous pruning methods, our method optimizes not only the parameters of the model but also the structure of the model. We train NRUPU in RTX2080Ti and deploy tests on edge devices NVIDIA Jetson Xavier NX (NJXN) and Atlas 200 DK (A2DK). Extensive experiments are conducted on the dataset TuSimple, CULane and our belt dataset with 11,894 data. Results show that NRUPU achieves over 96% speed increase and over 66% parameter reduction on all datasets within 0.7% accuracy loss. The FPS can reach 749, 665 and 783 on RTX2080Ti, 133, 117 and 143 on NJNX, 178, 161 and 183 on A2DK respectively. The code is released at https://anonymous.4open.science/r/NRUPU-29B0.