To see a list of all previous events, check out this page.
- This event has passed.
Applied Mathematics Seminars: Remote Sensing and 3D Image Analysis
June 13 @ 3:00 pm - 4:30 pm
Speaker 1: Pereira Sánchez, Iván
Title: Referring Image Segmentation for Remote Sensing
Summary: This seminar presents preliminary work on applying Referring Image Segmentation (RIS) to remote sensing imagery. RIS is a multimodal task that segments specific objects in images based on natural language descriptions. We focus on adapting the Densely Connected Parameter-Efficient Tuning for Referring Image Segmentation (DETRIS) model to this domain. DETRIS leverages a Parameter-Efficient Tuning (PET) approach to adapt pretrained CLIP and DINOv2 encoders for textual and visual inputs, respectively. By introducing small trainable modules while keeping the original weights frozen, the model retains prior knowledge while adapting to remote sensing data.
Our study includes an ablation analysis of modifications across the model’s architecture. We compare different text encoders (CLIP vs. BERT), implement differential attention in each decoder layer, and explore a more complex design for the projection module that generates segmentation masks. These investigations aim to optimize RIS performance for remote sensing tasks.
Speaker 2: Sánchez Beeckman, Marco
Title: Three-dimensional image analysis for almond endocarp feature extraction and shape description
Summary: This seminar summarizes our recently published computational approach to characterizing almond endocarp morphology using 3D scanning and computer vision. The method extracts quantitative descriptors from 3D meshes of almond endocarps, enabling systematic analysis of shape diversity. Over 9500 endocarps were scanned, aligned via affine transformations, and processed to extract features like contour and apex shape, keel development, surface markings, and symmetry. Compared to 2D imaging, 3D computational analysis provides richer geometric descriptions while still reducing human biases and inconsistencies associated with traditional visual evaluations. The descriptors are validated on 2610 endocarps from 36 Mallorcan and 14 international almond varieties, demonstrating alignment with existing qualitative morphological classifications. The results suggest that 3D imaging offers a reliable, bias-resistant alternative to visual assessments in agrobiodiversity studies.