In this work, I introduce ComicsPAP, a large-scale benchmark for comic strip understanding with over 100k samples organized into 5 subtasks. Through this Pick-a-Panel framework, I evaluate state-of-the-art MLLMs and demonstrate their limitations in capturing sequential and contextual dependencies, while also proposing adaptations that achieve better performance than models 10x larger.

In this survey, I review comics understanding through the lens of vision-language models, introducing a new taxonomy, the Layer of Comics Understanding (LoCU), to redefine tasks in comics research, analyze datasets, and highlight challenges and future directions for AI applications in comics' unique visual-textual narratives.

In this paper, I propose GCPL and CoMPLe to improve fine-grained categorization in vision-language models using generative modeling and contrastive learning, outperforming existing few-shot recognition methods.

CoMix is a multi-task comic analysis benchmark covering object detection, character identification, and multi-modal reasoning, using diverse datasets to balance styles beyond manga. I evaluate models in zero-shot and fine-tuning settings, revealing a gap between human and model performance, and set a new standard for comprehensive comic understanding.

In this work, I introduce the Comics Datasets Framework, standardizing annotations and addressing manga overrepresentation with the Comics100 dataset. I benchmark detection architectures to tackle challenges like small datasets and inconsistent annotations, aiming to improve object detection and support more complex computational tasks in comics.

In this work, I introduce a Multimodal-LLM for a comics text-cloze task, improving accuracy by 10% over existing models through a domain-adapted visual encoder and new OCR annotations, and extend the task to a generative format.

In this paper, I propose a smart pen and deep learning-based approach for early dysgraphia detection in children, offering a faster and more objective alternative to the traditional BHK test, validated through handwriting samples and expert interviews.

In this paper, I introduce Contextualized Table Extraction (CTE), a task to extract and structure tables within their document context, supported by a new dataset of 75k annotated pages from scientific papers, combining data from PubTables-1M and PubLayNet.

In this paper, I introduce a framework for Multilingual Scene Text Visual Question Answering (MUST-VQA) that handles new languages in a zero-shot setting, evaluates models in IID and zero-shot scenarios, and demonstrates the effectiveness of adapting multilingual language models for the STVQA task.

In this work, I address the problem of table extraction in scientific papers using Graph Neural Networks with enriched representation embeddings to improve distinction between tables, cells, and headers, evaluated on a dataset combining PubLayNet and PubTables-1M.