JPV-Net: Joint Point-Voxel Representations for Accurate 3D Object Detection
Authors: Nan Song, Tianyuan Jiang, Jian Yao2271-2279
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on the KITTI dataset and the ONCE dataset demonstrate that our proposed JPV-Net outperforms other state-of-the-art methods with remarkable margins. |
| Researcher Affiliation | Academia | 1School of Remote Sensing and Information Engineering, Wuhan University, Wuhan, Hubei, P.R. China 2AI Application and Innovation Research Center, The Open University of Guangdong, Guangzhou, Guangdong, P.R. China {nanaoisong, tianyuan, jian.yao}@whu.edu.cn |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not explicitly state that open-source code for the described methodology is provided or linked. |
| Open Datasets | Yes | KITTI Dataset (Geiger, Lenz, and Urtasun 2012) is a widely used benchmark dataset for autonomous driving. ... Furthermore, we also evaluate our method on the larger and more diverse ONCE dataset, and experimental results are illustrated in the supplementary file. (Mao et al. 2021) |
| Dataset Splits | Yes | The training samples are generally divided into the train split (3,712 samples) and the val split (3,769 samples). ... our models are trained on the train split and evaluated on the val split for validation, while trained with all training samples and evaluated with test samples for test. |
| Hardware Specification | Yes | Our JPV-Net framework is end-to-end trainable by the ADAM optimizer with an initial learning rate and a weight decay of 0.01, and the batchsize 16 on 8 GTX 1080 Ti GPUs. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies like libraries or frameworks, only mentioning optimizers and loss functions. |
| Experiment Setup | Yes | Our JPV-Net framework is end-to-end trainable by the ADAM optimizer with an initial learning rate and a weight decay of 0.01, and the batchsize 16 on 8 GTX 1080 Ti GPUs. We train our network for 80 epochs with the cosine annealing strategy for the learning rate decay. In training, the Io U thresholds for positive and negative anchors are set to 0.6 and 0.45, respectively. ... For inference, we only keep the top-100 proposals for further refinement. After the refinement, redundant boxes are removed with a NMS threshold of 0.1. |