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..
GSAlign: Geometric and Semantic Alignment Network for Aerial-Ground Person Re-Identification
Authors: Qiao Li, Jie Li, Yukang Zhang, Lei Tan, Jing Chen, Jiayi Ji
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | A comprehensive evaluation on CARGO with four matching protocols demonstrates the effectiveness of GSAlign, achieving significant improvements of +18.8% in m AP and +16.8% in Rank-1 accuracy over previous state-of-the-art methods on the aerial-ground setting. |
| Researcher Affiliation | Academia | Qiao Li1, Jie Li2 , Yukang Zhang2, Lei Tan3 , Jing Chen1, Jiayi Ji2,3 1Key Laboratory of Aerospace Information Security and Trusted Computing, Ministry of Education, School of Cyber Science and Engineering, Wuhan University 2Xiamen University 3National University of Singapore EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology using text and mathematical equations, accompanied by architectural diagrams (e.g., Figure 2). However, there are no explicit 'Pseudocode' or 'Algorithm' blocks or figures present in the document. |
| Open Source Code | Yes | The code is available at: https://github.com/stone96123/GSAlign. |
| Open Datasets | Yes | We conduct experiments on the CARGO dataset [8], a large-scale benchmark specifically designed for aerial-ground person re-identification (AG-Re ID). ...we also conduct comparative experiments on the AGRe ID [6] and AG-Re ID v2 [16] datasets... |
| Dataset Splits | Yes | Following the standard protocol, we use 51,451 images with 2500 identities for training and 51,024 images with the remaining 2,500 identities for testing. |
| Hardware Specification | No | The paper mentions 'FLOPs of 17.67 GFLOPs' and inference times (0.791 seconds per batch) but does not specify the exact GPU/CPU models or other detailed hardware specifications used for the experiments. |
| Software Dependencies | No | The paper mentions using 'Vi T-Base as the backbone' and 'optimized using Adam W', but it does not specify software versions for libraries like PyTorch, TensorFlow, Python, or CUDA, which are essential for reproducibility. |
| Experiment Setup | Yes | Input images are resized to 256 128, and horizontal flipping is applied during training for data augmentation. The model is optimized using Adam W with weight decay of 0.05 and a cosine learning rate schedule. The initial learning rate is set to 3.5 10 4 with linear warm-up for the first 20 epochs. λ is set to 0.1 during the training. We train for 120 epochs in total, using a batch size of 64. |