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..
FastJAM: a Fast Joint Alignment Model for Images
Authors: Omri Hirsch, Ron A Shapira Weber, Shira Ifergane, Oren Freifeld
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on several benchmarks demonstrate that Fast JAM achieves results better than existing modern JA methods in terms of alignment quality, while reducing computation time from hours or minutes to mere seconds. |
| Researcher Affiliation | Academia | Omri Hirsch Ron Shapira Weber Shira Ifergane Oren Freifeld The Faculty of Computer and Information Science, Ben Gurion University of the Negev (BGU), Israel The Data Science Research Center, BGU The School of Brain Sciences and Cognition, BGU EMAIL EMAIL |
| Pseudocode | No | The paper describes the method narratively and through mathematical equations (e.g., equations 4, 5, 6) and figures (e.g., Figure 2 for architecture overview) but does not include a dedicated 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | Yes | Our code is available at our project webpage, https://bgu-cs-vil.github.io/Fast JAM/. |
| Open Datasets | Yes | We use two benchmark datasets: SPair-71k [9] and CUB-200 [67] (classes and subsets). SPair-71k s test set comprises 18 object categories, each with 30 images, with annotated KPs and large intra-class variation. We report both per-category performance and average results across all categories. Following prior works, we evaluate on the first 3 categories of CUB-200 test set, each containing 30 images as well. |
| Dataset Splits | Yes | We use two benchmark datasets: SPair-71k [9] and CUB-200 [67] (classes and subsets). SPair-71k s test set comprises 18 object categories, each with 30 images, with annotated KPs and large intra-class variation. We report both per-category performance and average results across all categories. Following prior works, we evaluate on the first 3 categories of CUB-200 test set, each containing 30 images as well. |
| Hardware Specification | Yes | All experiments were conducted on a single NVIDIA RTX 4090 GPU with 24GB of memory. |
| Software Dependencies | No | Our GNN implementation is based on torch_geometric, primarily using the Graph SAGE [4] architecture, along with other variants for comparison. We used Weights & Biases for experiment tracking and visualization. For keypoint matching, we built upon the official implementation of Ro Ma [5], and we also incorporated components from the Lo FTR framework [6]. Object-centric masks were obtained using Grounded-SAM, which combines Grounding DINO [7, 8] and the Segment Anything Model (SAM) [9]. |
| Experiment Setup | Yes | We optimize Fast JAM for 600 epochs using Adam [66] with a Geman Mc Clure robustness parameter σ = 0.25. We use pretrained Grounding-SAM [57] and Ro Ma [13] with the default HP once, before starting the optimization. We train the model for 600 epochs using the Adam optimizer with an initial learning rate of 5 10 3, multiplied by 0.5 after 200 epochs without improvement. The loss function is based on the Geman Mc Clure formulation with a robustness parameter σ = 0.25, and no weight decay is applied. The feature extractor within the GNN uses 5 layers of hidden size 128, followed by a linear projection to an 8-dimensional homography parameter vector. ... During optimization, horizontal flips are checked every 100 epochs. |