We design several models to evaluate papers from different perspectives, including appearance, text coherence, structure and element diversity.
-
Deep Paper Gestalt
-
Gradient Boosting Decision Tree
-
Paper-sequence-based LSTM
-
Attention-based RCNN
CVPR 2015 - 2020 (you can use crawler.py to get these papers from CVFs on your own.)
| Methods | Accuracy | F1-score |
|---|---|---|
| Paper-image+ResNet18 | 84% | 76% |
| Lightgbm | 87% | 81% |
| Paper-seq+CNN (Ours) | 86.84% | 77.96% |
| Paper-seq+CNN LSTM (Ours) | 89.41% | 83.18% |
| Paper-seq+LSTM (Ours) | 90.30% | 84.59% |
| Attention-based RCNN (Ours) | 88% | 81% |