MGNet: Learning Correspondences via Multiple Graphs
Authors: Dai Luanyuan, Xiaoyu Du, Hanwang Zhang, Jinhui Tang
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that MGNet outperforms state-of-the-art methods in different visual tasks. The code is provided in https://github.com/DAILUANYUAN/MGNet-2024AAAI. Main Datasets. Yahoo s YFCC100M (Thomee et al. 2016) and SUN3D (Xiao, Owens, and Torralba 2013) datasets are chosen as outdoor and indoor scenes, respectively. Main Evaluation Metrics. The error metrics can be defined by angular differences between calculated rotation/translation vectors (recovered from the essential matrix) and the ground truth. m AP5 and m AP20 are selected as the default metrics in the camera pose estimation task. |
| Researcher Affiliation | Collaboration | Luanyuan Dai1, Xiaoyu Du1, Hanwang Zhang2, Jinhui Tang1* 1Nanjing University of Science and Technology, China 2Nanyang Technological University, Singapore {dailuanyuan, duxy, jinhuitang}@njust.edu.cn, hanwangzhang@ntu.edu.sg |
| Pseudocode | No | No pseudocode or clearly labeled algorithm blocks are present in the paper. |
| Open Source Code | Yes | The code is provided in https://github.com/DAILUANYUAN/MGNet-2024AAAI. |
| Open Datasets | Yes | Main Datasets. Yahoo s YFCC100M (Thomee et al. 2016) and SUN3D (Xiao, Owens, and Torralba 2013) datasets are chosen as outdoor and indoor scenes, respectively. |
| Dataset Splits | Yes | Following OA-Net++ (Zhang et al. 2019), 68 sequences are selected as training sequences and the rest 4 sequences are regarded as unknown scenes in outdoor scenes, and 239 sequences are chosen as training sequences, and the rest 15 sequences are unknown scenes in indoor scenes. Incidentally, we divide training sequences into three parts, consisting of training (60%), validation(20%) and testing (20%), and the last one is used as known scenes. |
| Hardware Specification | Yes | Experiments are performed on NVIDIA GTX 3090 GPUs. |
| Software Dependencies | No | The paper mentions "Adam (Paszke et al. 2017) optimizer is used" but does not specify version numbers for any software dependencies like Python, PyTorch, or specific libraries. |
| Experiment Setup | Yes | Network input is N 4 initial correspondences by SIFT or Super Point, and typically N is up to 2000. Cluster number m, neighbor number k and channel dimension S are 100, 24 and 128. Batchsize and β in Equation 14 are set to 32 and 0.5, respectively. Adam (Paszke et al. 2017) optimizer is used with a learning rate of 10 3 and we choose a warmup strategy. Clearly, a linearly growing rate is used for the first 10k iterations, after that the learning rate begins to decrease and reduce for every 20k iterations with a factor of 0.4. |