Improving Transformer Based Line Segment Detection with Matched Predicting and Re-ranking
Authors: Xin Tong, Shi Peng, Baojie Tian, Yufei Guo, Xuhui Huang, Zhe Ma
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our method outperforms other Transformer-based and CNN-based approaches in prediction accuracy while requiring fewer training epochs than previous Transformer-based models. |
| Researcher Affiliation | Collaboration | Xin Tong, Shi Peng, Baojie Tian, Yufei Guo, Xuhui Huang, Zhe Ma* Intelligent Science & Technology Academy of CASIC EMAIL, EMAIL |
| Pseudocode | No | The paper describes methods and formulas but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any statement or link indicating the availability of open-source code for the described methodology. |
| Open Datasets | Yes | We conduct our experiments in two publicly available datasets including the Wireframe dataset (Huang et al. 2018) and the York Urban dataset (Denis, Elder, and Estrada 2008). |
| Dataset Splits | Yes | The Wireframe dataset consists of 5,000 training and 462 test images of man-made environments, while the York Urban dataset has 102 test images. Following the typical training and test protocol (Huang et al. 2020b; Zhou, Qi, and Ma 2019), we train our model with the training set from the Wireframe dataset and test with both Wireframe and York Urban datasets. |
| Hardware Specification | Yes | We use 4 NVIDIA Tesla V100 cards for training and 1 card for evaluation. |
| Software Dependencies | No | Our training and evaluation are implemented in PyTorch. We use AdamW as the model optimizer and set weight decay as 10 4. While PyTorch is mentioned, no specific version number is provided. |
| Experiment Setup | Yes | We use AdamW as the model optimizer and set weight decay as 10 4. We train the model for 120 epochs. The initial learning rates are set to 5 10 4 for all parameters. Learning rates are reduced by a factor of 10 in epoch 60 and 90. We use a batch size of 8 and the size of the input images is set to 512 512. ... λr, λc, λp, λj, λe are set to 1, 1, 10, 1, 1, respectively. m is set as 32 for sampling and δe, δd, δl are set to 0.5, 0.5, 0.5 empirically for YUD dataset and 0.4, 0.1, 0.2 after validing on random sampling set on training set. 500 line segments with high scores are detected with NMS for the proposed method in all the experiences. |