Notice: This website is continuously updated as new conference proceedings and journal volumes are published. Consequently, the figures presented here may differ from those reported in our publications.
A brief overview of our methodology is provided on the Methods page, along with references to publications documenting our approach and links to the code and data used in each study.
The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates.
RAPTR: Radar-based 3D Pose Estimation using Transformer
Authors: Sorachi Kato, Ryoma Yataka, Pu (Perry) Wang, Pedro Miraldo, Takuya Fujihashi, Petros Boufounos
NeurIPS 2025 | Venue PDF | LLM Run Details | Input Tokens: 21,193 Total number of tokens sent to the LLM as input for this paper's analysis. | Output Tokens: 4,617 Total number of tokens produced by the LLM (including reasoning/thinking tokens) for this paper's analysis.
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Evaluated on two indoor radar datasets, RAPTR outperforms existing methods, reducing joint position error by 34.3% on HIBER and 76.9% on MMVR. Table 1 shows the performance of 3D pose estimation for HIBER. Table 2 shows the performance comparison for baselines and RAPTR with MMVR. 6 Ablation Study |
| Researcher Affiliation | Collaboration | 1Mitsubishi Electric Research Laboratories (MERL), USA 2The University of Osaka, Japan 3Information Technology R&D Center (ITC), Mitsubishi Electric Corporation, Japan |
| Pseudocode | No | The paper describes the architecture and methodology in sections like "4 RAPTR: Radar-based 3D Pose Estimation using Transformer" with figures and mathematical equations, but does not present any explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our implementation is available at https://github.com/merlresearch/radar-pose-transformer. |
| Open Datasets | Yes | Datasets: We assess the performance of RAPTR and baseline models on the HIBER dataset3 [35] and the MMVR dataset4 [24], both of which are publicly available multi-view mm Wave radar datasets designed for indoor human perception tasks. 3https://github.com/Intelligent-Perception-Lab/HIBER 4https://zenodo.org/records/12611978 |
| Dataset Splits | Yes | The HIBER dataset includes two-view radar heatmaps from 10 different viewpoints, the corresponding 3D keypoint labels, and the 3D BBox labels. We use data protocols MULTI and WALK , and use views 2 through 10 for training, validation, and testing. The MMVR dataset includes two-view radar heatmaps in various indoor scenarios, the corresponding 2D keypoint labels, and the 3D BBoxes. We use a data split P1S1 , a single-person case in an open space. |
| Hardware Specification | Yes | Justification: Refer Table 7 in Appendix F including descriptions for computer resources we used. Hardware RTX A6000 |
| Software Dependencies | Yes | Justification: We included the specifications of data and hyper parameters. Refer Table 7 in Appendix F. Software PyTorch 1.12.1 CUDA 11.3 |
| Experiment Setup | Yes | Parameter Settings for RAPTR: We use T = 4 consecutive frames as input to our RAPTR network. For the point decoder, the number of pose queries N is 10. For the joint decoder, the number of joint queries K depends on the dataset to be evaluated: K = 14 for HIBER and K = 17 for MMVR. The parameters relating to model training are summarized in Appendix F. Batch size 2 Learning Rate 1e-4 Optimizer AdamW Epochs 100 Weight decay 1e-4 LR decay 1e-1 at 80, 90 epoch |