Dustin Carrión-Ojeda

I'm a PhD student at the Image and Video Analysis Group (IVA) at TU Darmstadt, under the supervision of Prof. Dr. Sc. Simone Schaub-Meyer. My work lies at the intersection of computer vision, limited-data learning, and multimodal representation learning, with a strong emphasis on efficiency and robustness.

I received my MSc in Artificial Intelligence from Université Paris-Saclay, where I worked with Prof. Isabelle Guyon, and my BEng in Information Technology from Yachay Tech University, where I was advised by Prof. Israel Pineda and Prof. Rigoberto Fonseca-Delgado.

Email  /  CV  /  Bio  /  Google Scholar  /  GitHub  /  Twitter  /  LinkedIn

Research

My research investigates how computer vision models can learn effectively under limited supervision and heterogeneous modality settings. In particular, I develop meta-learning approaches for image classification and segmentation, and design multimodal methods for video understanding that remain robust under missing or incomplete modalities.

A central theme of my work is efficiency, reducing data requirements, memory footprint, and computational cost while maintaining strong generalization performance. I also have prior experience in EEG-based biometric systems, exploring the intersection of biological signal processing and machine learning.

Publications

A multimodal knowledge distillation framework for egocentric action recognition that remains robust under missing modalities, without requiring modality alignment during training or inference. The distilled student leverages available modalities while reducing computational cost by approximately 50%.

An efficient transformer that processes high-resolution correlation tokens, boosting few-shot classification and segmentation, especially for small objects, while using at least four times fewer parameters than existing methods. It supports N-way K-shot tasks and correctly outputs empty masks when no target is present.

An analysis of EEG-based biometrics identifying the most relevant features and channels, reducing required electrodes from 32 to 11 while maintaining performance, with standard deviation from wavelet coefficients as the best feature.

A domain-specific language built on Keras, designed for fast and accessible neural network development with minimal coding.

An analysis of the NeurIPS'22 Cross-Domain MetaDL Challenge results, emphasizing the impact of pre-trained backbones, overfitting prevention, and domain adaptation techniques in improving meta-learning performance across 10 domains.

A diverse meta-dataset designed for few-shot image classification and other applications, featuring 40 curated datasets from 10 domains, standardized formatting, and multiple dataset versions.

A challenge on "any-way any-shot" image classification in a cross-domain setting. Winning solutions were blind-tested and open-sourced for broader impact.

An automated system for issuing fines for speed limit violations using vehicular networks, featuring a dissemination protocol that achieves approximately 95% fine delivery in urban scenarios.

A user-friendly Python framework for developing EEG-based applications, enabling in-depth analysis of factors that influence the performance of the system and its computational cost.

An analysis of factors influencing the performance of EEG-based biometrics, showing that EEG recording time significantly impacts classifier accuracy.

A novel EEG-based biometric system that uses Discrete Wavelet Transform (DWT) for feature extraction and achieves approximately 90% accuracy with just 2 seconds of EEG recording.

Posters

Based on the design and source code from Jon Barron's website