AD-PT: Autonomous Driving Pre-Training with Large-scale Point Cloud Dataset

Authors: Jiakang Yuan, Bo Zhang, Xiangchao Yan, Botian Shi, Tao Chen, Yikang LI, Yu Qiao

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental By conducting extensive experiments, we argue that there are two key issues that need to be solved for achieving the real AD-PT... Our study provides a more unified approach, meaning that once the pre-trained checkpoint is generated, it can be directly loaded into multiple perception baselines and benchmarks. Results further verify that such an AD-PT paradigm achieves large accuracy gains on different benchmarks (e.g., 3.41%, 8.45%, 4.25% on Waymo, nu Scenes, and KITTI).
Researcher Affiliation Collaboration Jiakang Yuan1, , Bo Zhang2, , Xiangchao Yan2, Tao Chen1, , Botian Shi2, Yikang Li2, Yu Qiao2 1School of Information Science and Technology, Fudan University 2Shanghai Artificial Intelligence Laboratory
Pseudocode No The paper describes the model architecture and mathematical formulations but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes Project page: https://jiakangyuan.github.io/AD-PT.github.io/. Our code is based on 3DTrans [22].
Open Datasets Yes Pre-training Dataset. ONCE [14] is a large-scale dataset... Description of Downstream Datasets. 1) Waymo Open Dataset [20] contains 150k frames. 2) nu Scenes Dataset [1] provides point cloud data... 3) KITTI Dataset [5] includes 7481 training samples.
Dataset Splits Yes The labeled set divides into a train set with 6 sequences ( 5K samples), a validation set with 4 sequences ( 3k frames), and a test set with 10 sequences ( 8k frames)... Waymo Open Dataset [20]... divided into a train set with 798 sequences ( 150k samples) and a validation set with 202 sequences ( 40k samples)... nu Scenes Dataset [1]... consisting of 28130 training samples and 6019 validation samples. KITTI Dataset [5] includes 7481 training samples and is divided into a train set with 3712 samples and a validation set with 3769 samples.
Hardware Specification Yes We pre-train for 30 epochs on ONCE small split and large split using 8 NVIDIA Tesla A100 GPUs.
Software Dependencies No We use Adam optimizer with one-cycle learning rate schedule... Our code is based on 3DTrans [22]. The paper does not provide specific version numbers for software dependencies like PyTorch, Python, or the 3DTrans codebase.
Experiment Setup Yes We use Adam optimizer with one-cycle learning rate schedule and the maximum learning rate is set to 0.003. We pre-train for 30 epochs on ONCE small split and large split... The number of selected Ro I features is set to 256 and the cross-view matching threshold τ is set to 0.3.