STONE: A Submodular Optimization Framework for Active 3D Object Detection

Authors: RUIYU MAO, Sarthak Kumar Maharana, Rishabh Iyer, Yunhui Guo

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our method achieves state-of-the-art performance with high computational efficiency compared to existing active learning methods. We extensively validate the proposed framework on real-world autonomous driving datasets, including KITTI [15] and Waymo Open dataset [52], achieving state-of-the-art performance in active 3D object detection.
Researcher Affiliation Academia Ruiyu Mao Sarthak Kumar Maharana Rishabh K Iyer Yunhui Guo Department of Computer Science The University of Texas at Dallas, Richardson, TX, USA {rxm210041, skm200005, rishabh.iyer, yunhui.guo}@utdallas.edu
Pseudocode No The paper describes its algorithms in prose (e.g., 'A greedy algorithm is then applied...') and provides an illustrative pipeline in Figure 1, but it does not include a formal pseudocode or algorithm block.
Open Source Code Yes The code is available at https://github.com/RuiyuM/STONE
Open Datasets Yes For our experiments, we use the KITTI dataset [15], one of the commonly used datasets in autonomous driving tasks. ... We also use the more challenging dataset for 3D object detection in autonomous driving the Waymo Open dataset [52].
Dataset Splits Yes The dataset consists of 3,712 training samples and 3,769 validation samples, which include a total of 80,256 labeled objects. These objects include cars, pedestrians, and cyclists, each annotated with class categories and bounding boxes. We also use the more challenging dataset for 3D object detection in autonomous driving the Waymo Open dataset [52]. It includes 158,361 training samples and 40,077 testing samples.
Hardware Specification Yes We train our proposed method and all baselines on a GPU cluster with 4 NVIDIA RTX A5000 GPUs.
Software Dependencies No The paper states, 'We adopt the implementation settings as outlined in CRB.' and discusses the backbone model (PV-RCNN), but it does not explicitly list specific version numbers for general software dependencies like Python, PyTorch, or CUDA in the main text.
Experiment Setup Yes For KITTI and Waymo Open, the training batch sizes are set to 6 and 4 respectively. However, the evaluation batch sizes are set to 16 for both datasets. We optimize the network parameters using Adam with a fixed learning rate of 0.01. For all the methods, we perform 5 stochastic forward passes of the MC-Dropout [14]. Active learning parameters. For KITTI, we set Γ1 and Γ2 to 400 and 300 respectively. In the case of Waymo Open, Γ1 and Γ2 are set to 2,000 and 1,200 respectively. To ensure fairness, Nq is set to 100 in all the methods.