Dual-Perspective Knowledge Enrichment for Semi-supervised 3D Object Detection

Authors: Yucheng Han, Na Zhao, Weiling Chen, Keng Teck Ma, Hanwang Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on the two benchmark datasets demonstrate that our DPKE achieves superior performance over existing state-of-the-art approaches under various label ratio conditions.
Researcher Affiliation Collaboration Yucheng Han1, Na Zhao2*, Weiling Chen3, Keng Teck Ma3, Hanwang Zhang1 1Nanyang Technological University 2Singapore University of Technology and Design 3Hyundai Motor Group Innovation Center in Singapore
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The source code will be made available to the public.
Open Datasets Yes We validate our proposed method on the Scan Net (Dai et al. 2017) and SUN RGB-D (Song, Lichtenberg, and Xiao 2015) datasets.
Dataset Splits Yes We conduct experiments with label ratios of 0.05, 0.1, 0.2, and 1.0, as presented in Table 1. For both the Scan Net and SUN RGB-D datasets, we follow the standard preprocessing (Zhao, Chua, and Lee 2020; Wang et al. 2021a) and evaluation protocols to ensure consistency across different experiments. These protocols involve reporting the mean average precision (m AP) over three data splits with 3D Intersection over Union (Io U) thresholds of 0.25 and 0.5.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running experiments.
Software Dependencies No The paper states "Our code is built based on (Wang et al. 2021a)" but does not list specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes Our code is built based on (Wang et al. 2021a), adhering to their optimization and training parameters. The student and teacher models are initialized with parameters exclusively trained on labeled data during the pretraining stage. In the training stage, the student model s weights are updated using the loss function gradients, while the teacher model s weights are updated using the exponential moving average of the student model s weights. Following previous works (Wang et al. 2021a; Wu et al. 2022; Zhao, Chua, and Lee 2020), our DPKE utilizes different augmentation methods for the student and teacher models. The teacher model undergoes weak augmentation of sub-sampling on the input point clouds. On the other hand, the student model is subjected to a strong augmentation scheme, including random flipping and scaling of the point clouds. We conduct experiments with label ratios of 0.05, 0.1, 0.2, and 1.0, as presented in Table 1.