Elamur Research⌕
☰
home / papers / 2026 / No Overfit at NakbaArchiveClassifier Shared Task: A Swin Transformer-Based System for Destruction Image Classification

No Overfit at NakbaArchiveClassifier Shared Task: A Swin Transformer-Based System for Destruction Image Classification

Paper · 2026 · Mohamed Fathy Mohamed, Samar Mahmoud Abd El-Mageed and Ensaf Mohamed

Summary

The paper reports a team's entry in the NakbaArchiveClassifier shared task (Nakba-NLP 2026 @ LREC 2026). The task is binary classification of photos as showing damaged or destroyed infrastructure versus intact infrastructure. The official data consists of 2,001 Instagram images shared by Gaza-based journalists and content creators since October 2023, and only 1,400 of them are available for training. To cope with so little labelled data, the authors used a Swin Transformer (tiny variant) pretrained on ImageNet-1K. They froze its first three stages and trained only a head made of the final Swin stage plus an MLP classifier. They also grew the training set to 10,000 images by pseudo-labelling images from external disaster and damage datasets, then manually checking the predictions the model was unsure about. On the blind CodaBench test set, the system reached 86.55% F1 and 87.81% accuracy and placed 5th. The work shows that a frozen pretrained backbone combined with semi-supervised data expansion is a practical recipe for conflict-documentation archives with few labels.

Key points

  • Architecture: The frozen backbone uses Swin stages with 2, 2 and 6 blocks, expanding channels from 96 to 384. The trainable head has a further Swin stage (768 dimensions), global average pooling and an MLP with two 256-unit layers, GELU activation, dropout and BCE loss.
  • Training schedule: The MLP was trained alone for 30 epochs, then the whole head for 70 more epochs. The optimizer was AdamW (weight decay 1e-4), with learning rates of 1e-4 and then 1e-5 and a batch size of 32. Training ran on a single Tesla P100 on Kaggle.
  • Data expansion:
    • About 8,600 images were added from four sources: Tuncer (2023), Niloy et al. (2021), a war-related image set, and the Alam et al. (2020) disaster-response benchmark.
    • Predictions with confidence above 0.9 or below 0.1 were kept as pseudo-labels, and all other images were re-annotated by hand.
  • Test results: F1 86.55, accuracy 87.81, precision 86.88, recall 86.26, ranking 5th.
  • Ablation (F1):
    • Swin + MLP: 86.55
    • EfficientNet: 84.56
    • Swin without an extra head: 84.51
    • CNN + Swin hybrid: 83.56
    • Swin + SVM: 82.03

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Nile UniversityGiza · Egypt14 works · 158 citations🎓Helwan UniversityCairo · Egypt8 works · 36 citations

Also on the byline: Luxor University.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksComputer Vision and Pattern Recognition · 2,869 worksEngineering · 133 worksMedia Technology · 34 worksCivil and Structural Engineering · 2 works

Cite

Cite as Mohamed Fathy Mohamed, Samar Mahmoud Abd El-Mageed and Ensaf Mohamed. 2026. No Overfit at NakbaArchiveClassifier Shared Task: A Swin Transformer-Based System for Destruction Image Classification. Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/28o7zedpovh7

@inproceedings{mohamed2026no,
  title = {{No Overfit at NakbaArchiveClassifier Shared Task: A Swin Transformer-Based System for Destruction Image Classification}},
  author = {Mohamed Fathy Mohamed, Samar Mahmoud Abd El-Mageed and Ensaf Mohamed},
  booktitle = {Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026},
  year = {2026},
  doi = {10.63317/28o7zedpovh7},
  url = {https://doi.org/10.63317/28o7zedpovh7}
}

References

15 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. 2025. stable dataset. https://universe.roboflow.com/nothing-cniqf/ stable-d0jqc. Roboflow Dataset. ↗
  2. Alexei Abrahams, Shadi Abudalfa, Mustafa Jarrar, and George Mikros. 2026. The nakbaarchiveclassifier shared task on nakba image classification. In Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources (Nakba-NLP 2026), co-located with the Language Resources and Eval… find
  3. Firoj Alam, Ferda Ofli, and Muhammad Imran. 2018. Crisismmd: Multimodal twitter datasets from natural disasters. In Proceedings of the 12th International AAAI Conference on Web and Social Media (ICWSM). find
  4. Firoj Alam, Ferda Ofli, Muhammad Imran, Tanvirul Alam, and Umair Qazi. 2020. Deep learning benchmarks and datasets for social media image classification for disaster response. In 2020 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pages 151–158. find
  5. Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee. find
  6. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An image is worth 16x16 words: Transformers for image recognition at scale. In Intern…
  7. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778.
  8. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, volume 25, pages 1097–1105.
  9. Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012– 10022.
  10. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR). find
  11. Hussein Mozannar, Yara Rizk, and Mariette Awad. 2018. Damage identification in social media posts using multimodal deep learning. find
  12. Fahim Faisal Niloy, Abu Bakar Siddik Nayem, Anis Sarker, Ovi Paul, M Ashraful Amin, Amin Ahsan Ali, Moinul Islam Zaber, AKM Mahbubur Rahman, et al. 2021. A novel disaster image data-set and characteristics analysis using attention model. In 2020 25th International Conference on Pattern Recognition (… find
  13. Karen Simonyan and Andrew Zisserman. 2015. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations (ICLR). Published as a conference paper at ICLR 2015. Originally submitted in 2014.
  14. Mingxing Tan and Quoc Le. 2019. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning (ICML), pages 6105–6114.
  15. Türker Tuncer. 2023. Damaged constructions image dataset. https://www.kaggle.com/datasets/turkertuncer/ damaged-constructions-image-dataset. Kaggle dataset. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mohamed-2026-no-overfit-nakbaarchiveclassifier-shared-task/