Notice: 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. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
SingRef6D: Monocular Novel Object Pose Estimation with a Single RGB Reference
Authors: Jiahui Wang, Haiyue Zhu, Haoren Guo, Abdullah Al Mamun, Cheng Xiang, Tong Heng LEE
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Evaluations of pose estimation on the REAL275, Clear Pose, and Toyota-Light datasets show that our approach surpasses state-of-the-art methods, achieving a 6.1% improvement in average recall. Project page: https://plusgrey.github.io/singref6d. |
| Researcher Affiliation | Academia | Jiahui Wang1 Haiyue Zhu2 Haoren Guo1 Abdullah Al Mamun1 Cheng Xiang1 Tong Heng Lee1 1College of Design and Engineering, National University of Singapore 2SIMTech, Agency for Science, Technology and Research (A*STAR) |
| Pseudocode | No | The paper describes its methodology using descriptive text, mathematical formulas, and pipeline diagrams, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The code, for privacy reasons, is not included in the initial submission stage. Upon the situation of acceptance, we will consider releasing the code. |
| Open Datasets | Yes | Evaluations of pose estimation on the REAL275, Clear Pose, and Toyota-Light datasets show that our approach surpasses state-of-the-art methods... REAL275 [61] is chosen for its complex scenes with various objects. Toyota-Light [62], which is included in the BOP [62] challenge suite, provides a standardized testbed for evaluating robustness under challenging lighting conditions.To validate our method on transparent objects, we include Clear Pose [13] and downsample the whole training set with a step size of 100. |
| Dataset Splits | Yes | For the REAL275 [61]and Tyo-Light [62] datasets, we randomly select 80% of the separated scenes as fine-tuning data, and the rest of the scenes are processed as testing data. The Clear Pose [13] dataset was constructed using sets 1, 4, 5, 6, and 7. Scene 5 from set 1 and scene 6 from sets 4-7 were designated as the test set due to their distinguished visual environments, while the remaining samples were randomly allocated to the training and testing sets. ... The processed clearpose dataset contains 3129 images for training and 643 images for testing. |
| Hardware Specification | Yes | All of our experiments are conducted on an Ubuntu 22.04 server with two Nvidia RTX3090 GPUs. |
| Software Dependencies | Yes | The deep learning framework is Py Torch 2.2.0 with CUDA 12.4. |
| Experiment Setup | Yes | We fine-tune our model for 80 epochs with a batch size of 16 for all datasets except Clear Pose [13], which is fine-tuned with 40 epochs for a smaller time consumption. We adopt an Adam W optimizer with an initial learning rate of 0.0005, updated by a cosine annealing scheduler with 3 warmup epochs. To avoid overfitting, we assign a weight decay of 0.01 to the optimizer. ... In this section, we illustrate the detailed hyper-parameter settings for our fine-tuning process in Table 10. |