Pedestrian-Centric 3D Pre-collision Pose and Shape Estimation from Dashcam Perspective
Authors: MeiJun Wang, Yu Meng, Zhongwei Qiu, Chao Zheng, Yan Xu, Pengxiaorui , Jian Gao
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our framework achieved promising results on the PVCP dataset, outperforming other methods of human pose estimation. and 5 Experimental and Results |
| Researcher Affiliation | Collaboration | Meijun Wang1, Yu Meng1 , Zhongwei Qiu2, Chao Zheng1, Yan Xu1, Xiaorui Peng1, Jian Gao3 1University of Science and Technology Beijing 2Alibaba DAMO Academy 3Northwest University |
| Pseudocode | No | The paper describes the network architecture and methods but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code and data are available for research at https://github.com/wmj142326/PVCP. |
| Open Datasets | Yes | Code and data are available for research at https://github.com/wmj142326/PVCP. and while the pre-training model weights were obtained from Motion BERT (12) trained on the AMASS (17), Human3.6M (16), PW3D (21), and MSCOCO (20) datasets. |
| Dataset Splits | No | Subsequently, we selected 164 video sequences as the trainset and 45 video sequences as the testset. |
| Hardware Specification | Yes | four NVIDIA RTX 2080Ti GPUs for all training |
| Software Dependencies | No | Py Torch (65) was used for the entire experimental environment |
| Experiment Setup | Yes | batchsize was uniformly set to 32. In the training stage, we only use the images from the PVCP trainset and the corresponding 2D ground truth keypoints as the input of the two models. We first train the ITP network by loading a pre-trained model of the MPII dataset (51) and training 40 epoches. For PTM, we use sequence length T = 16 and train 100 epochs in about two hours. |