Multimodal Image Matching Based on Cross-Modality Completion Pre-training
Authors: Meng Yang, Fan Fan, Jun Huang, Yong Ma, Xiaoguang Mei, Zhanchuan Cai, Jiayi Ma
IJCAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments demonstrate that XCP-Match outperforms existing algorithms on public datasets. Section 4 is dedicated to "Experiments" covering various evaluations and an ablation study. |
| Researcher Affiliation | Academia | 1Wuhan University 2Macau University of Science and Technology EMAIL, EMAIL. All affiliations are universities, and email domains are academic (.edu.cn, .edu.mo). |
| Pseudocode | No | The paper describes the methodology in narrative text and uses figures to illustrate the architecture (e.g., Figure 1 and Figure 2). It does not contain any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any statement about releasing source code or a link to a code repository. The text is ambiguous and lacks a clear, affirmative statement of release. |
| Open Datasets | Yes | In pre-training, we use the KAIST Multispectral Pedestrian dataset [Hwang et al., 2015] for training. In fine-tuning, we use the Mega Depth dataset [Li and Snavely, 2018]. To evaluate the performance of XCP-Match for relative pose estimation in visible-infrared image pairs, we test it on the METU-Vis TIR dataset [Tuzcuo glu et al., 2024]. To test XCP-Match for homography estimation, we use the Road Scene dataset [Xu et al., 2020b; Xu et al., 2020a]. To test XCP-Match s performance for image registration, we evaluate it on the Tri Modal Human dataset [Palmero et al., 2016]. |
| Dataset Splits | No | The paper mentions using datasets for training and testing (e.g., "KAIST Multispectral Pedestrian dataset [...] for training" and "Mega Depth dataset [...] for fine-tuning"), and it mentions selecting specific images for testing in one case ("We use RGB-FIR images and select only those with distinct human segmentations for testing"). However, it does not provide specific information on how the datasets are split into training, validation, and test sets (e.g., percentages, sample counts, or references to predefined splits with detailed methodology) to reproduce the data partitioning. |
| Hardware Specification | Yes | Pre-training is conducted using the Adam W optimizer with a learning rate of 2.5 10 4, a batch size of 2, a total of 15 epochs, and 30 hours of training on 2 NVIDIA Ge Force RTX 4090 GPUs. Fine-tuning is conducted [...] on 2 NVIDIA Ge Force RTX 4090 GPUs. |
| Software Dependencies | No | The paper mentions using the "Adam W optimizer" but does not specify any programming languages, libraries, or frameworks with version numbers that would be necessary to replicate the experiment. |
| Experiment Setup | Yes | Pre-training is conducted using the Adam W optimizer with a learning rate of 2.5 10 4, a batch size of 2, a total of 15 epochs, and 30 hours of training on 2 NVIDIA Ge Force RTX 4090 GPUs. Fine-tuning is conducted [...] with a learning rate of 2.5 10 4, a batch size of 2, a total of 25 epochs, and 125 hours of training on 2 NVIDIA Ge Force RTX 4090 GPUs. The thresholds in the matching network are set to: θc = 0.3, θf = 0.1. The settings in the loss function are set to: λc = 0.5, λf = 0.3, λsub = 104, λs = 1, λvis ac = λir ac = 0.25. |