Use TensorRT network definition API and plugin to deploy 3D object detection algorithm SE-SSD

Published: 07 Apr 2023, Last Modified: 10 Apr 2023ICLR 2023 Workshop SR4AD HYBRIDReaders: Everyone
TL;DR: https://github.com/jingyue202205/SE-SSD-AI-TRT
Abstract: SE-SSD: Self-Ensembling Single-Stage Object Detector From Point Cloud (CVPR 2021) [[Paper\]](https://arxiv.org/abs/2104.09804) The usual way to deploy a deep learning model to tensorRT is to convert it to onnx, and use tensorRT onnxLoad API for inference. However, for simple models, this method is feasible, but for complex models, the model cannot be directly converted to onnx, so it is more difficult to deploy. We found the github repository of tensorRTx(https://github.com/wang-xinyu/tensorrtx), which focuses on deploying complex image-based models to tensorRT. Although this repo provides a basic solution. It is still difficult to deploy SE-SSD on tensorRT using this method. The main difficulty is that the SE-SSD algorithm includes 3D Submanifold sparse convolution and 3D sparse convolution(implemented by libtorch, c++, cuda). These two convolutions are difficult to implement with tensorRT plugin because they contain parameters such as trainable weights. We spent a lot of time trying to solve complex model deployment problems, trying to write weighted operations as tensorRT plugins. In the end, we successfully deployed the SE-SSD algorithm to tensorRT(https://github.com/jingyue202205/SE-SSD-AI-TRT). Although the current running speed is relatively slow, we will optimize the running speed later.
Track: Original Contribution
Type: Repository
0 Replies

Loading