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..
SegMASt3R: Geometry Grounded Segment Matching
Authors: Rohit Jayanti, Swayam Agrawal, Vansh Garg, Siddharth Tourani, Muhammad Haris Khan, Sourav Garg, Madhava Krishna
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that our approach outperforms state-of-the-art methods, including the SAM2 video propagator and local feature matching methods, by up to 30% on the AUPRC metric, on Scan Net++ and Replica datasets. |
| Researcher Affiliation | Academia | Rohit Jayanti1 Swayam Agrawal1 Vansh Garg1 Siddharth Tourani2,3 Muhammad Haris Khan3 Sourav Garg4 Madhava Krishna1 1IIIT Hyderabad 2University of Heidelberg 3MBZUAI 4Independent |
| Pseudocode | No | The full algorithm is provided in the supplementary. (Referring to LFM's algorithm, not the paper's core method in the main text). Figure 1 provides an overview of our method. It builds upon the MASt3R [22] architecture by introducing a Feat2Seg Adapter... |
| Open Source Code | No | We will release code and image pairs trained on upon paper acceptance. |
| Open Datasets | Yes | Our network is trained on scenes from Scan Net++ [47] which contain a diverse set of real-world scenes... We test our model on novel scenes from Scan Net++ as well as perform cross-dataset generalization studies on Replica [39] and Map Free [2] datasets. |
| Dataset Splits | Yes | We train on 860 k image pairs from 140 scenes and evaluate on 8 k pairs from 36 validation scenes, sampled with a fixed seed (42) and balanced across scenes and four pose bins: [0 45 ], [45 90 ], [90 135 ], and [135 180 ], defined by the rotational geodesic distance between camera orientations. |
| Hardware Specification | Yes | We use a batch size of 36 and train the model for 20 epochs on a single NVIDIA RTX A6000 GPU. |
| Software Dependencies | No | The model is trained using Adam W optimizer with an initial learning rate of 1e-4, weight decay of 1e-4, and a cosine annealing learning rate schedule without restarts, decaying up to a minimum learning rate of 1e-6 over the full training duration. |
| Experiment Setup | Yes | The model is trained using Adam W optimizer with an initial learning rate of 1e-4, weight decay of 1e-4, and a cosine annealing learning rate schedule without restarts, decaying up to a minimum learning rate of 1e-6 over the full training duration. We use a batch size of 36 and train the model for 20 epochs on a single NVIDIA RTX A6000 GPU. |