Pasquale De Marinis | University of Bari Aldo Moro

About me

I am a Ph.D. student in Computer Science at the Computational Intelligence Lab (CILAB), University of Bari Aldo Moro.

My research focuses on data-efficient and explainable deep learning for computer vision, with applications in large-scale and real-world visual understanding.

I am currently working on few-shot semantic segmentation and edge AI methods developed in collaboration with the Jheronimus Academy of Data Science (Netherlands).

Research interests

Computer Vision: Semantic segmentation, Scene understanding, Object detection
Machine Learning: Few-shot learning, Transfer learning, Explainable AI
Applied AI: Edge intelligence, Efficient deep architectures, Sustainable computing

News

[September 2025] I chaired the workshop “Advances in Drone Vision” at ICIAP 2025.
[August 2025] We received the “Outstanding Paper Award” at IFSA-NAFIPS 2025 for our paper “Explainable Fuzzy GNNs for Leak Detection in Water Distribution Networks”.
[July 2025] Our Paper “Label Anything: Multi-Class Few-Shot Semantic Segmentation with Visual Prompts” has been accepted to ECAI 2025.
[May 2025] We are organizing the workshop “International Workshop on Advances in Drone Vision” at ICIAP 2025.
[September 2024] I started my visit at the Jheronimus Academy of Data Science (JADS) in ‘s-Hertogenbosch, Netherlands.
[August 2024] Our Paper “RoWeeder: Unsupervised Weed Mapping through Crop-Row Detection” has been accepted to CVPPA at ECCV 2024.
[October 2023] Our Paper “Weed mapping in multispectral drone imagery using lightweight vision transformers” has been accepted to Neurocomputing.
[September 2023] I received the Best Distinction Paper Award at FEDCSIS 2023.
[July 2023] I participated to DeepLearn Summer School 2023 in Gran Canaria, Spain.
[April 2023] I participated as student volunteer to DeepLearn Spring 2023 in Bari, Italy.

Featured publications

ECAI2025

Label Anything: Multi-Class Few-Shot Semantic Segmentation with Visual Prompts

Pasquale De Marinis*, Nicola Fanelli, Raffaele Scaringi, Emanuele Colonna, Giuseppe Fiameni, Gennaro Vessio, Giovanna Castellano

European Conference on Artificial Intelligence (ECAI), 2025.

PDF Code Project Page

We introduce Label Anything, a transformer-based framework for multi-prompt, multi-way few-shot semantic segmentation. Our model unifies diverse visual prompts—points, boxes, and masks—into a single, flexible architecture that generalizes across tasks while reducing annotation effort. Evaluated on the COCO-20i benchmark, it achieves state-of-the-art performance and demonstrates strong multi-class segmentation capabilities.

ECCVW2024

RoWeeder: Unsupervised Weed Mapping through Crop-Row Detection

Pasquale De Marinis*, Gennaro Vessio, Giovanna Castellano

European Conference on Computer Vision Workshops (ECCVW), 2024.

PDF Code Project Page

We present RoWeeder, an unsupervised framework for weed mapping that integrates crop-row detection with a noise-resilient deep learning model. By generating pseudo-ground truth from crop-row geometry, our lightweight network learns to distinguish crops from weeds under noisy conditions. Evaluated on the WeedMap dataset, RoWeeder surpasses existing baselines and enables real-time, drone-based weed monitoring for precision agriculture.

Neurocomputing

Weed mapping in multispectral drone imagery using lightweight vision transformers

Giovanna Castellano, Pasquale De Marinis*, Gennaro Vessio

Neurocomputing

PDF Code Project Page

We introduce a lightweight Vision Transformer for drone-based multispectral weed mapping in precision agriculture. The model transfers knowledge from RGB pretraining to multispectral data, achieving state-of-the-art results on the WeedMap dataset and enabling efficient, non-invasive weed detection for targeted herbicide management.

FEDCSIS

Applying Knowledge Distillation to Improve Weed Mapping With Drones

Giovanna Castellano, Pasquale De Marinis*, Gennaro Vessio

18th Conference on Computer Science and Intelligence Systems

PDF Code Project Page Best Distinction Paper Award

Our approach employs state-of-the-art knowledge distillation to transfer performance from a large teacher network to a compact student model, achieving high accuracy on the WeedMap dataset with a fraction of the computation. The method enables efficient, drone-deployable weed detection for sustainable precision agriculture.