Francis Engelmann’s Post

Machine learning in real-time from a smartphone lens! Real-time scan to render to search is transformative for all types of experiences from Sports to Surveillance. NerfStudio is known for providing a simplified end-to-end process for creating, training, and testing Neural Radiance Fields (NeRFs). It supports a more interpretable implementation by modularizing each component, aiming to enhance user experience in exploring NeRF technology. Ready to take the AI journey together?

Dear Colleagues, The FRCSyn: Face Recognition Challenge in the Era of Synthetic Data will be held on January 8 from 8:20am to 12:45pm HST - Hawaii Standard Time. The workshop will be conducted in a virtual format at WACV 2024. You can find the detailed schedule of the FRCSyn Workshop at https://frcsyn.github.io/ and the summary paper of the FRCSyn Challenge at https://lnkd.in/di_X57Wh We have the pleasure of featuring the following keynote speeches during the workshop: 1) "In the Era of Synthetic Avatars: From Training to Verification" by Koki Nagano, NVIDIA Research. 2) "Zero-shot/few-shot learning for model diagnosis and debiasing generative models, and its applications to face analysis" by Fernando De La Torre, Carnagie Mellon University. 3) "Biometric Recognition in the Era of AI Generated Content (AIGC)" by Xiaoming Liu, Michigan State University. We look forward to your participation. The FRCSyn Organizers: Pietro Melzi, Minchul Kim, Ruben Tolosana, Christian Rathgeb, Ruben Vera Rodriguez, Aythami Morales Moreno, Xiaoming Liu, Julian FIERREZ, Javier Ortega-Garcia. #WACV #syntheticdata #facerecognition #computervision #biometrics #machinelearning #privacy #bias #research #challenge #FRCSyn #artificialintelligence #IEEE #CVF

Robotic manipulation of deformable objects, such as clothes and fabric, is widely recognized as a complex challenge, particularly from the perspectives of perception and decision-making. The inherent unpredictability of the environment's dynamics adds to the complexity, making it a important research area for applying statistical learning techniques and neural policies. Hugging Face leRobot library, combined with the low-cost-robot architecture, is breaking down barriers to entry in this research field and providing open access to datasets, learning algorithms, and policies which enables greater reproducibility and collaboration in robot learning research.

#EuropeanAI LSTMs (developed by Sepp Hochreiter) in the form of the #xLSTM architecture are back! And outperform transformers (the architecture behind OpenAI) for certain computer vision tasks! Shout-Out to ELLIS - European Laboratory for Learning and Intelligent Systems (founded by Bernhard Schölkopf et al)! A new model developed by Maximilian Beck and Benedikt Alkin, named "Vision-LSTM" (#ViL) (link in the comments) uses #xLSTM for image processing - instead of transformers. Early tests show it works very well - eg outperforming other models on ImageNet-1K (a benchmarking data set used by AI researchers). And because #xLSTM scales linearly with sequence length instead of quadratically, it could overcome computing limitations (example in the comments)! So #ViL is a strong alternative for eg. medical imaging tasks. This is very interesting, because the #xLSTM architecture was originally designed for understanding language - and not even for computer vision. If you are an AI startup, looking for funding, please reach out to us at DTCF | DeepTech & Climate Fonds

Excited to present my latest project showcasing fascinating models from Hugging Face, utilizing huggingface.co and Google Colab under the guidance of #Saisatesh at AIMER Society - Artificial Intelligence Medical and Engineering Researchers Society. Hugging Face is revolutionizing Al with community-driven model development. I imported a model from Hugging Face and configured a Colab environment with a T4 GPU. Successfully ran the model on various images, achieving accurate detections and outputs. Topics covered in this video include: 1. Summarization 2. Fill-Mask 3. Visual Question Answering #AIMERS #AIMERSOCIETY #APSCHE #DATASCIENCE #OPENCV #HUGGINGFACE #ARTIFICIALINTELLIGENCE #MACHINELEARNING #DEEPLEARNING #BIGDATA #COMPUTERVISION #NATURALLANGUAGEPROCESSING

Estimating camera parameters from a set of images depicting a scene is a crucial task in computer vision. The traditional feature-based structure-from-motion (SfM) tools like COLMAP have been solving this task by incremental reconstruction. However, the authors of a recent paper from Niantic Labs re-interpreted this approach and investigated alternative visual relocalizers that are not rooted in local feature matching. They introduced ACE0 (ACE Zero), a novel method that estimates camera poses with an accuracy comparable to feature-based SfM, without requiring pose priors nor sequential inputs. ACE0 builds implicit, neural scene representations from unposed images, optimizing efficiently over thousands of images. Check out the project, arxiv, video, and code at the links below! Project: https://lnkd.in/eZiB3k-X Arxiv: https://lnkd.in/ePHjnRvU Video: https://lnkd.in/exmE8yUK Code: https://lnkd.in/ey2Wh9qc

I'm thrilled to announce that we have two papers accepted for presentation at IEEE/CVF WACV 2024. 🔹INCODE: Implicit Neural Conditioning with Prior Knowledge Embeddings Our first paper, INCODE, introduces a new method that enhances Implicit Neural Representations (INRs) by leveraging deep prior knowledge. INCODE consists of a harmonizer network and a composer network, where the harmonizer network dynamically adjusts critical parameters of the composer's activation function. This leads to remarkable improvements in signal representation across various domains, including audio, image, and 3D shape. It also excels in addressing complex challenges such as neural radiance fields (NeRFs) and solving inverse problems like denoising, super-resolution, inpainting, and CT reconstruction. 📃 Project Page: https://lnkd.in/er9gXKfy 📚 Arxiv: https://lnkd.in/e6DBzEJs 💻 GitHub: https://lnkd.in/eCHefthw 🔹Beyond Self-Attention: Deformable Large Kernel Attention for Medical Image Segmentation Our second paper introduces "Deformable Large Kernel Attention (D-LKA Attention)," a novel approach designed to elevate medical image segmentation. This method efficiently captures volumetric context using large convolution kernels, avoiding excessive computational demands. D-LKA Attention also benefits from deformable convolutions to adapt to diverse data patterns. We have created both 2D and 3D versions, with the 3D version excelling in cross-depth data understanding. This forms the basis of our new hierarchical Vision Transformer architecture, the D-LKA Net, which outperforms existing methods on popular medical segmentation datasets (Synapse, NIH Pancreas, and Skin lesion). 📚 Arxiv: https://lnkd.in/ezq66XCy 💻 GitHub: https://lnkd.in/eu42DJRW I would like to express my sincere gratitude to Reza Azad, Alireza Hosseini, Ehsan Khodapanah Aghdam, Leon Niggemeier, Prof. Ulaş Bağcı, Prof. Merhof, and others for their invaluable contributions. #WACV2024 #ComputerVision #MedicalImageSegmentation #AIResearch #INR

Hello Connections, I recently completed a simple yet exciting project involving object detection using a webcam and OpenCV with YOLOv8. This powerful model can detect virtually anything with remarkable accuracy. By leveraging diverse datasets, we can achieve expressive and reliable object detection results. For this project, I utilized Roboflow, a fantastic platform for managing datasets. Specifically, I worked with a dataset of UNO cards, but many other datasets can be used to explore different detection tasks. I have learned a great deal from Sai Satish Sir while interning at AIMER Society - Artificial Intelligence Medical and Engineering Researchers Society. The experience has been invaluable and has significantly enhanced my skills in AI and object detection. #aimers #apsche #aiml #yolov8 #opencv #objectdetection #cnnproject #unocarddataset

🚀 Thrilled to announce that I have successfully completed Task 3: Object Detection using YOLO, with the invaluable support of AIMER Society - Artificial Intelligence Medical and Engineering Researchers Society!🚀 This journey has equipped me with cutting-edge skills in object detection, enabling me to develop and deploy models that can identify and classify objects in real-time. A heartfelt thank you to Aimer Society for their continuous guidance and encouragement. Eager to leverage these new capabilities in upcoming projects and further my expertise in the field of machine learning and computer vision. 🌟 #ObjectDetection #YOLO #MachineLearning #ComputerVision #AimerSociety #ContinuousLearning

You've likely marveled at the lifelike physics in video games, where wind rustles through trees and waves crash against the shore, all thanks to complex mathematical modeling. But what if I told you about something even more groundbreaking? OpenAI's Sawa is changing the game, not by simulating physics through traditional methods, but by fundamentally understanding it. Imagine a system that doesn't rely on intensive computational models to replicate the dance of leaves in the wind or the water flow around a rock—it just knows. This leap towards an intuitive AI understanding of fluid dynamics and physical interactions is impressive; it's a glimpse into a future where technology perceives the world with a depth akin to human knowledge. #OpenAI #FutureOfAI #InnovativeTechnology