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..
BMW: Bidirectionally Memory bank reWriting for Unsupervised Person Re-Identification
Authors: Xiaobin Liu, Jianing Li, Baiwei Guo, WenbinZhu, Jing Yuan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on standard benchmarks demonstrate the effectiveness of our BMW method in unsupervised Re ID model training. Specially, BMW even outperforms previous methods that use stronger backbones. Code is available at https://github.com/liu-xb/BMW. |
| Researcher Affiliation | Collaboration | Xiaobin Liu Nankai University Tianjin, China EMAIL Jianing Li Hong Kong Polytechnic University Hong Kong, China EMAIL Baiwei Guo Huawei Technologies Ltd. Shanghai, China EMAIL |
| Pseudocode | No | The paper describes the method using mathematical equations and textual explanations of steps, but it does not include explicitly labeled pseudocode or algorithm blocks. For example, Section 4 details "Bidirectionally Memroy bank re Writing" with equations like (2), (3), (4), (5), (6), (7), (8), (9), (10), and (11). |
| Open Source Code | Yes | Code is available at https://github.com/liu-xb/BMW. |
| Open Datasets | Yes | Experiments are performed on Market-1501 [54] and MSMT17 [27]. Market-1501 contains 32,668 images of 1,501 identities captured from 6 cameras at Tsinghua University. MSMT17 contains 126,441 images of 4,101 identities captured from 15 cameras at Peking University. Due to the data usage restriction, experiments on Duke MTMC-re ID [55] are provided on the repository on github. Limited by the paper length, experimental results on the vehicle dataset Ve Ri-776 [60] are provided in Sec. A.5. |
| Dataset Splits | Yes | Market-1501 contains 32,668 images of 1,501 identities captured from 6 cameras at Tsinghua University. 12,936 images of 751 identities are selected for training, and others are used for testing. In the test set, 3,368 images are selected as query images and remaining 19,732 images are used as gallery images. MSMT17 contains 126,441 images of 4,101 identities captured from 15 cameras at Peking University. 32,621 images of 1,041 identities are selected for training, and others are selected for testing. In the test set, 11,659 images are selected as query images and remaining 82,161 images are used as gallery images. Ve Ri-776 contains 37,746 images of 576 vehicles for training, 1,678 images and 51,003 images of another 200 vehicles are used as queries and galleries, respectively. |
| Hardware Specification | Yes | Model is trained on a server with 4 RTX 4090 GPU, and 128G memory. |
| Software Dependencies | No | The paper mentions using DBSCAN, Adam optimizer, ResNet50, and ImageNet as components or pre-trained models. However, it does not specify version numbers for any general software libraries, frameworks (e.g., PyTorch, TensorFlow), or specific solvers required to replicate the experiments. |
| Experiment Setup | Yes | Parameter τ in Eqn. (1) is set to 0.05 following [22]. λintra, and λinter are set as 0.9, and 0.02, respectively. More details are provided in Sec. A.4. ... DBSCAN is used for the unsupervised clustering and the eps is set as 0.6 following [22]. Input images are resized to 256 128. We use random flipping, random cropping, and random erasing [69] for data augmentation. The Adam optimizer is adopted for training. Learning rate is initialized as 0.00035 and decayed by 0.1 every 25 epochs. Model is totally trained for 75 epochs. In each training batch, we randomly sample 16 identities, i.e., clusters by DBSCAN. For each sampled identity, 16 images are randomly selected to compose the training batch, resulting 256 images. |