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 [1].
SSAN: A Symbol Spatial-Aware Network for Handwritten Mathematical Expression Recognition
Authors: Haoran Zhang, Xiangdong Su, Xingxiang Zhou, Guanglai Gao
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that SSAN significantly improves the recognition performance of the HMER models, and the proposed auxiliary tasks are more effective in enhancing HMER performance than existing auxiliary tasks. ... To evaluate the effectiveness of the proposed method, we conduct the experiment on the benchmark datasets named CROHME (Mouch ere et al. 2014) and HME100K (Yuan et al. 2022). We use the Expression Recognition Rate (Exp Rate), 1 Error, and 2 Error as the metrics to evaluate the performance of different methods in HMER. |
| Researcher Affiliation | Academia | Haoran Zhang1,2,3, Xiangdong Su1,2,3 *, Xingxiang Zhou1,2,3, Guanglai Gao1,2,3 1College of Computer Science, Inner Mongolia University, China 2National & Local Joint Engineering Research Center of Intelligent Information Processing Technology for Mongolian, China 3Inner Mongolia Key Laboratory of Multilingual Artificial Intelligence Technology, China EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes methods and processes using textual descriptions and architectural diagrams (e.g., Figure 2, Figure 3), but it does not contain a specific block labeled as "Pseudocode" or "Algorithm" with structured steps. |
| Open Source Code | Yes | Code https://github.com/Howrunz/SSAN |
| Open Datasets | Yes | To evaluate the effectiveness of the proposed method, we conduct the experiment on the benchmark datasets named CROHME (Mouch ere et al. 2014) and HME100K (Yuan et al. 2022). |
| Dataset Splits | No | The paper mentions using "CROHME (Mouch ere et al. 2014) and HME100K (Yuan et al. 2022)" as benchmark datasets, but it does not explicitly state the training, validation, or test splits (e.g., percentages or exact counts) used for these datasets. It refers to 'CROHME 2014', 'CROHME 2016', 'CROHME 2019' which are specific versions of the CROHME benchmark. |
| Hardware Specification | Yes | All models are trained on two NVIDIA V100 32GB GPUs, and the batch size is 20. For a fair comparison, we provide results with and without data augmentation. We employ scale augmentation (Li et al. 2020) with a scaling factor from 0.7 to 1.4 in the experiment with data augmentation. ... The FLOPs and FPS were calculated on the HME100K test dataset using an NVIDIA Tesla V100 GPU. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software libraries, programming languages, or frameworks used (e.g., Python, PyTorch, TensorFlow, CUDA versions). |
| Experiment Setup | Yes | The output feature channels for the two SSA submodules are 600 and 432. The kernel size and stride of the max-pooling layer are both set to 2. All models are trained on two NVIDIA V100 32GB GPUs, and the batch size is 20. ... In the first stage, we jointly optimize the HMER model with the SSAN. Here, the learning target of SSAN is the symbol spatial distribution map of the printed template corresponding to the input formula images. Through joint optimization with SSAN in the first stage, the decoder of the HMER model has been able to learn better attention maps than without SSAN. In spite of this, it is worth noting that there is loose alignment between the symbol spatial distribution maps of the handwritten mathematical expressions and that of their printed templates. To ensure the self-consistency of the HMER model (Farquhar et al. 2021; Bonatti and Mohr 2022; Wang et al. 2023), we adjust the learning target of SSAN in the second stage of training to the attention graph of the trained HMER decoder in the first stage. |