Viviane Zulian’s Post

🔍 Just published a new blog post on the challenges of Out-Of-Distribution (OOD) robustness in open-vocabulary object detection models. This study compares the zero-shot capabilities of OWL-ViT, YOLO World, and Grounding DINO under distribution shifts. Check it out here: https://bit.ly/3UPI4JK

New RSLab paper titled "Multiannual Change Detection in Long and Dense Satellite Image Time Series Based on Dynamic Time Warping". Link to the paper: https://lnkd.in/d-bdE3H7 #RemoteSensing #EarthObservation #ChangeDetection #TimeSeries #LandCoverChange

Samed's work, 'Impact of Surface Reflections in Maritime Obstacle Detection', has been accepted to the BMVC 2024 Workshop on Robust Recognition in the Open World: https://lnkd.in/da_ce8Vu Pre-print version of the paper: https://lnkd.in/dxyhgYVZ Datasets: https://lnkd.in/dtNNXeVZ In this work, we show that surface reflections indeed adversely affect detector performance. We measure the effect of reflections by testing on two custom datasets, which we make publicly available. The first one contains imagery with reflections, while in the second reflections are inpainted. We show that the reflections reduce mAP by 1.2 to 9.6 points across various detectors. To remove false positives on reflections, we propose a novel filtering approach named Heatmap Based Sliding Filter. We show that the proposed method reduces the total number of false positives by 34.64% while minimally affecting true positives. Congratulations Samed Yalçın!

Have you ever looked for a dataset for your own classifier but didn't find the correct data or quantity? Well, why don't you just generate it yourself? That is what I did on my recent research. Where we generated a dataset for a remote sensing classifier which performed very well on classifying real remote sensing images, knowing that it only trained on generated remote sensing images. Shout-out to my co-authors Abdullah Al Salmani and Ashwaq Alkaabi. Read the full paper: https://lnkd.in/dmvHWZBK https://lnkd.in/dvbFJ5te

Day 2 of #30DayMapChallenge: In Germany, nationwide GTFS files are generally published weekly. These files are collected from multiple data providers, including transport associations, across the country and made available as a centralized nationwide dataset. In 2022, 49 files were published. A comparison with the central stop directory shows that the number of stops with assigned trips varies by region from week to week. The map shows the number of GTFS datasets in which a 5 km geogrid cell is likely to contain fewer stops in the GTFS dataset than actually exist. This map was part of a publication earlier this year with Holger Bruch.

We implement and experiment on the beta-Variational Autoencoder framework introduced by Higgins et al. published in ICLR in 2017 TLDR: The idea is we tweak the vanilla Variational Autoencoder (VAE) by adding a coefficient to the Kullback-Leibler term of the ELBO loss function in order to give us more control of the latent representations, in particular, the separability of the representations across the latent space or latent variable’s dimensions. Setting Beta to 0 leads to more accurate input reconstructions but less "expressive" outputs when asked to generate from a randomly sampled latent space vector. Setting Beta to a very high number causes the model to lose reconstruction capabilities and give the same plausible but uninformative outputs. Pls. check my blog post for details https://lnkd.in/gdhyzwZn

A Year of Journal Articles (Day 256/365): Shi, Yonggang, Shaokun Li, Ziyan Liu, Zhiguo Zhou, and Xuehua Zhou. "MTP-YOLO: You Only Look Once Based Maritime Tiny Person Detector for Emergency Rescue." Journal of Marine Science and Engineering 12, no. 4 (2024): 669. Summary: Goal: Develop a highly accurate detector for tiny people in maritime environments (critical for emergency rescue). Challenges: People appear very small in vast sea backgrounds, making detection difficult. Proposed Method: MTP-YOLO ("You Only Look Once" based Maritime Tiny Person detector). Key components: Cross-stage partial layer with C2fELAN (preserves key features during calculations). Multi-level Cascaded Enhanced Convolutional Block Attention Module (MCE-CBAM) - focuses on object location. Weighted Efficient Intersection over Union (W-EIoU) Loss function (improves position and scale regression). Benefits: High accuracy for tiny person detection. Low number of parameters (efficient for real-world applications). Validation: Evaluated on the TinyPersonv2 dataset with promising results. Applications: Maritime emergency rescue missions.

🌊 Dive into oceanographic insights with Galaxy! Join FAIR-EASE on April 12th, 2024, from 11:00 to 12:00 (Online) for a webinar on "Harnessing Galaxy for Oceanographic Insights with ODV and DIVAnd Interpolation." Marie Jossé will guide you through using Galaxy for oceanographic data analysis, focusing on ODV and DIVAnd. Learn how to manage ODV tools, visualize, subset, and save data, and use DIVAnd in a JupyterLab environment with notebooks. Register now: https://lnkd.in/eccSxrzs #UseGalaxy #Oceanography #DataAnalysis #Webinar

It was an honor to host Akhil Nekkanti from Caltech at our data-driven physical simulation (#DDPS) seminar, Lawrence Livermore National Laboratory, on September 27, 2024. He talked about "data-driven techniques for analysis of turbulent flows." You can watch his talk through https://lnkd.in/gthdj6V2 He gave a wonderful talk, so please take advantage of it. #Abstract: #Turbulent #flows are high-dimensional systems characterized by instabilities and non-linearity, which make modeling challenging. Data-driven techniques reduce complexity by extracting key flow features and projecting governing equations onto a low-dimensional subspace. Recently, #spectral #proper #orthogonal #decomposition (#SPOD), a frequency-domain variant of principal component analysis, has emerged as a powerful tool for analyzing turbulent flows. We extend SPOD to include low-rank reconstruction, denoising, and frequency-time analysis. In this talk, I will demonstrate two applications: #gappy-#data #reconstruction and the intermittency of coherent structures. First, our gappy-data reconstruction algorithm uses spatial and temporal correlations to estimate compromised or missing regions, outperforming standard techniques like gappy POD and Kriging. Second, we introduce a convolution-based strategy for frequency-time analysis that characterizes the intermittency of spatially coherent flow structures. When applied to turbulent jet data, SPOD-based frequency-time analysis reveals that the intermittent occurrence of large-scale coherent structures is directly associated with high-energy events. Finally, we present #bispectral #mode #decomposition (#BMD), a technique that extracts flow structures linked to nonlinear triadic interactions by optimizing third-order statistics. This method is applied to a forced turbulent jet to examine and construct the cascade of triads. #DDPS #seminar is organized by #libROM team at LLNL. Visit https://www.librom.net to find more about libROM team. Information about #DDPS #seminar can be found at https://lnkd.in/gnGsx6qq

Three topics I #cannot #live #without: 🫶💜 (1) Turbulent flows 🌪️ (2) Dimensionality reduction 🤖 (3) Coherent structure identification 🕸️ #OrderFromChaos #DetailsMatter #ButNotEveryFeatureIsImportant 🙌👊