SASA: Semantics-Augmented Set Abstraction for Point-Based 3D Object Detection

Authors: Chen Chen, Zhe Chen, Jing Zhang, Dacheng Tao221-229

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on the popular KITTI and nu Scenes datasets validate the superiority of SASA, lifting point-based detection models to reach comparable performance to state-of-the-art voxel-based methods.
Researcher Affiliation Collaboration Chen Chen1, Zhe Chen1, Jing Zhang1, Dacheng Tao2,1 1 The University of Sydney, Australia 2 JD Explore Academy, China
Pseudocode Yes Algorithm 1: Semantics-guided Farthest Point Sampling Algorithm.
Open Source Code Yes Code is available at https://github.com/blakechen97/SASA.
Open Datasets Yes KITTI dataset (Geiger, Lenz, and Urtasun 2012) is a prevailing benchmark for 3D object detection in transportation scenarios.
Dataset Splits Yes Following the commonly applied setting (Zhou and Tuzel 2018), we divide all training examples into the train split (3, 712 samples) and the val split (3, 769 samples) and all experimental models are trained on the train split and tested on the val split.
Hardware Specification Yes The total batch size is set to 16, equally distributed on four NVIDIA V100 GPUs.
Software Dependencies No Our experimental models are all built with the Open PCDet (Team 2020) toolbox, including our reproduced 3DSSD and the official implementation of Point RCNN.
Experiment Setup Yes We train the 3DSSD model with ADAM optimizer for 80 epochs. We apply the one-cycle learning rate schedule (Smith and Topin 2019) with the peak learning rate at 0.01. The total batch size is set to 16, equally distributed on four NVIDIA V100 GPUs. During the training phase, manifold data augmentation strategies are employed to avoid over-fitting.