Improving Graph Matching with Positional Reconstruction Encoder-Decoder Network
Authors: Yixiao Zhou, Ruiqi Jia, Hongxiang Lin, Hefeng Quan, Yumeng Zhao, Xiaoqing Lyu
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental results on four public keypoint matching datasets demonstrate the effectiveness of our proposed PREGM. We evaluate our PREGM on four public keypoint matching datasets. Our experimental results demonstrate that PREGM outperforms state-of-the-art methods. We also present an ablation study, which shows the effectiveness of each component of PREGM. |
| Researcher Affiliation | Academia | Yixiao Zhou Wangxuan Institute of Computer Technology Peking University Ruiqi Jia Wangxuan Institute of Computer Technology Peking University Hongxiang Lin Wangxuan Institute of Computer Technology Peking University Hefeng Quan School of computer science and engineering Nanjing University of Technology Yumeng Zhao School of Artificial Intelligence Beijing University of Posts and Telecommunications Xiaoqing Lyu Wangxuan Institute of Computer Technology, Beijing Institute of Big Data Research Peking University |
| Pseudocode | No | No pseudocode or algorithm blocks were explicitly labeled or presented in a structured format. |
| Open Source Code | No | Our codes will be available in the future. |
| Open Datasets | Yes | We evaluate our PREGM on four public keypoint matching datasets: Pascal VOC [8], Willow Object Class [4], SPair-71k [23], and IMC-PT-Sparse GM [14]. The Pascal VOC dataset includes 20 classes of keypoints with Berkeley annotations [2] and images with bounding boxes. The Willow Object Class dataset contains images of five categories: face, duck, winebottle, car, and motorbike, the first three categories are from the Caltech-256 dataset [12], and the last two categories are from the Pascal VOC 2007 dataset [8]. The SPair-71k dataset is a relatively novel dataset, which was recently published in the paper [23] about dense image matching. Spair-71k contains 70958 image pairs of 18 categories from Pascal VOC 2012 dataset [8] and Pascal 3D+ dataset [41]. The IMC-PT-Sparse GM dataset contains 16 object categories and 25061 images [14], which gather from 16 tourist attractions around the world. |
| Dataset Splits | Yes | When conducting experiments on the Pascal VOC dataset, we follow the standard protocol [32, 37]: First, each object is cropped according to its corresponding bounding box and scaled to 256 256 px. Second, we use 7,020 images for training and 1,682 for testing. Following the default setting in [4, 32], we crop the images to the bounding boxes of the objects and rescale to 256 256 px, 20 images of each class are selected during training, and the rest are for testing. Following [29], we use 53,340 image pairs for training, 5,384 for validation, and 12,234 for testing, and we also scale each image to 256 256 px. We take 13 classes as the training set and the other 3 classes as the test set. Experiments are conducted on the benchmark with 50 anchors. |
| Hardware Specification | Yes | All experiments are run on a single GTX-1080Ti GPU. |
| Software Dependencies | No | The paper mentions using Adam [15] optimizer, vgg16_bn [30], Spline CNN [10], and LPMP solver [25], but no specific version numbers for these software components are provided. |
| Experiment Setup | Yes | We employ Adam [15] optimizer with an initial learning rate of 1 10 4 for PR-En Dec, and 9 10 4 for other models, and the learning rate is halved every three epochs. We empirically set lm = 3 and ln = 2 in the encoder and decoder, and choose temperature constant τ = 2 in Lcon, and balance factor λ = 1/32 in LP R En Dec. We also set batch size = 8 for Pascal VOC, Willow Object Class, and Spair-71k datasets. |