Valerio Marsocci’s Post

Object Detection Using Florence-2 🔥 The recently released Florence-2 model demonstrates strong zero-shot capabilities across tasks such as captioning, object detection, grounding, and segmentation. Below is an example of using the model for an object detection task and getting the bounding box coordinates for an image. The model is available on the Clarifai Platform. Try it out: https://lnkd.in/gxqa9nYp #objectdetection #computervision #vllm

Read how DUG Multi-parameter FWI Imaging breathes new life into vintage data, in this new feature on GEO EXPRO! https://lnkd.in/ge5cn2Jv

💥💥💥 Next-Token Prediction is All You Need Overview While next-token prediction is considered a promising path towards AGI, it has struggled to excel in multimodal tasks, which are still dominated by diffusion models (e.g., Stable Diffusion) and compositional approaches (e.g., CLIP combined with LLMs). In this work, we introduce Emu3, a new suite of state-of-the-art multimodal models trained solely with next-token prediction. By tokenizing images, text, and videos into a discrete space, we train a single transformer from scratch on a mixture of multimodal sequences. Project site 👉 https://lnkd.in/dAHQzZvQ #machinelearning Results 👇

This video depicts the difference in Surface Reconstruction quality between Photogrammetry, Neural Radiance Fields (NeRFs), 3D Gaussian Splats (SuGAR) and the latest 2D Gaussian Splats. I have been testing the code for different approaches and thought it would provide a visual reference as to where all of these technologies stand as of this moment. The dataset contains 27 images. I will write a more detailed comparison between the pipelines in a later post.

📢 #HighlyCited 1. Comparative Study Analysis of ANFIS and ANFIS-GA Models on Flow of Vehicles at Road Intersections https://lnkd.in/g5QGZFtR 2. Movement Time for Pointing Tasks in Real and Augmented Reality Environments https://lnkd.in/gNvhPSG9 3. A Novel Interval Iterative Multi-Thresholding Algorithm Based on Hybrid Spatial Filter and Region Growing for Medical Brain MR Images https://lnkd.in/gjbVjNMY 4. Multiscale Analysis for Improving Texture Classification https://lnkd.in/g8DWNkzF

A new study has been published concerning the determination of aerosol properties using images from an all-sky camera. The presented approach works in all sky conditions, but with focus partially cloudy scenes. A Gaussian Process Regression (GPR) machine learning has been used to associate RGB radiances with aerosol optical depth (AOD) and Angstrom exponent (AE) observed by AERONET. The method has been tested on a 2-year dataset with varying atmospheric conditions. AOD and AE predictions showed excellent agreement with AERONET measurements, particularly when using the quality assurance method.. Sensitivity analysis confirms the stability of our methodology, especially when using quality assurance criteria. This approach could potentially enhance the use of all-sky cameras in studying aerosol-cloud interactions in partially cloudy scenarios. https://lnkd.in/dbRSTzz4

Join my colleague Eugene Khvedchenia for a webinar tomorrow to learn how to boost accuracy and speed in satellite and aerial image object detection! Here's what you'll learn about: 🛰️ Discover YOLO-NAS-Sat and see how it stands up against the well-known YOLOv8 in small object detection tasks. 🏛️ A deep dive into YOLO-NAS-Sat's design, from backbone adjustments tailored for small objects to its innovative U-Net-style decoder that preserves critical details. 📈 Optimization Strategies: Gain actionable insights from Deci's experiences in fine-tuning models on the Dota 2.0 dataset, a vast collection of aerial images. 📊 Rethinking Evaluation Metrics: Investigate the shortcomings of traditional evaluation metrics like COCO mAP for small object detection and consider alternatives, including mAP@0.5 and distance-based metrics. Link to register in the comments! #computervision #deeplearning

🔥 #hottopic Multiscale Pixel-Level and Superpixel-Level Method for Hyperspectral Image Classification: Adaptive Attention and Parallel Multi-Hop Graph Convolution by Junru Yin, et al. ➡️ https://brnw.ch/21wLoTt

DEMONSTRATED that CNNs can fulfill most of the requirements for ocean remote sensing applications across the following six categories: reconstruction of the 3D ocean field, information extraction, image super resolution, ocean phenomena forecast, transfer learning method, and CNN model interpretability method.

#11 Seg-Eval Choosing the proper ground sampling distance is a pivotal decision in remote sensing downstream applications. In this work, the authors try to set out a clear ensemble of guidelines for fairly comparing semantic segmentation results obtained at various spatial resolutions. They also propose region-based pixel-wise metrics, allowing for a more detailed analysis of the model performance. what I liked: very practical and important problem paper: https://lnkd.in/gFfUE35v #50papers #AI4EO #remotesensing #deeplearning #GeoAI