Promaton’s Post

I'm extremely stoked to have contributed to this post on the Promaton blog - it was also a super fun and interesting project to work on, and it's great to have the opportunity to distil the knowledge into this simplified version we're now sharing with you! hope it's an enjoyable and insightful read - #computergraphics #react #threejs #dijkstra #algorithms #dentaldomain #blogpost

Image processing refers to performing operations on an image to enhance it or to extract useful information from it. Read our article explaining Image Processing operations and a simple program to detect faces in an image using Cascade classifier in skimage.feature module. #dataconfluence #facedetection #imageprocessing #skimage https://lnkd.in/d-MsmJpK

As an art analyst, recreating figures into their corresponding vector graphics is a fun job—until you are assigned a wordy figure. One of the most time-consuming parts of recreation is layout making for text items. Since I am not ready to use a beta version of my favorite image editor and am uncomfortable with online tools due to the confidential nature of the information, I tried some Python libraries and finally reached a solution: Automated OCR. Though I need to do more research on the 'confidentiality' part, I am pretty sure this works for me and serves as a perfect stress reducer. #science #scientificillustration https://lnkd.in/gN3jMSjr

Omost, a project aimed at enhancing AI-driven image generation. It converts large language models' (LLMs) coding capabilities into tools for creating and composing images. This project leverages multi-modal AI to make the process more efficient and effective. 🔹 Key Features: 1. LLM-Based Image Composition: Omost enables LLMs to write code for creating detailed visual content using a virtual Canvas. 2. Pretrained Models: We offer three pretrained models – omost-llama-3-8b, omost-dolphin-2.9-llama3-8b, and omost-phi-3-mini-128k. 3. Easy Deployment: Use the models directly on HuggingFace or deploy them locally (requires 8GB Nvidia VRAM). 🔹 Use Cases: Omost can generate various scenes, such as epic battles, post-apocalyptic escapes, gothic rulers, and humorous cartoons. 🔹 Technical Highlights: 1. Global and Local Descriptions: Define overall scenes and specific elements with detailed attributes. 2. Depth and Color Parameters: Enhance scene composition using distance_to_viewer and HTML_web_color_name. For more information, check out the Omost GitHub page. https://lnkd.in/gd3cVcAg

In many instances in medical imaging workflows, you'll need to go from a volumetric image to a 3D mesh. Here's how you can do this (and more!). There are several algorithms that can take a volumetric image and convert it into a mesh. The most widely known one is marching cubes or it's modern version, which is called flying edges. At PYCAD we do this operation quite often. We typically refer to it as NIFTI to STL conversion. I will share below, in the first comment, a python script that shows you how to do this. Moreover, we quite often need to do other types of conversions as well, such as: - DICOM to NIFTI or vise versa. - DICOM to NRRD or vise versa. - NIFTI to PNG. We have developed functions for all of these operations. You'll find them in the same folder that contains NIFTI to STL converter. What other conversions do you think we should implement next?

Excited to share my insights on mastering color spacing with y'all! So, color spaces in image processing are kind of a big deal. They're important for all sorts of apps and projects, especially when it comes to OpenCV. Now, one of the common tasks in this field is converting RGB (Red, Green, Blue) to BGR (Blue, Green, Red) format. Sounds simple, right? Well, it can actually make a big difference in your work, especially when it comes to tasks like image manipulation and computer vision. I've been digging deep into this whole RGB-to-BGR conversion thing lately, using NumPy and OpenCV to find the best ways to make it happen. And let me tell you, it's not as easy as it sounds! But, after lots of trial and error, I've come up with some pretty cool methods that make the conversion process a breeze, no matter what platform or app you're using. NumPy's array manipulation tools and OpenCV's image processing features are like the perfect partners in crime for this job. They make it super easy to switch those color channels around and make sure everything looks just right. Now, understanding color spaces isn't just about making your images look better. It also opens up all sorts of doors for creativity and innovation. You can experiment with different color schemes, create some seriously cool effects, and maybe even come up with some new ways to use image processing in the future. I'm stoked to share my findings with the awesome community of image processing and computer vision enthusiasts out there. Whether you're a seasoned pro or just getting started, mastering color spacing is a skill that can take your work to the next level. Let's geek out about color spaces together! #ImageProcessing #ComputerVision #OpenCV #NumPy #ColorSpaces #RGBtoBGR #DataScience #MachineLearning #TechCommunity

Take control over your AI image composition with Omost! Omost turns your prompt into instructions for LLM model to better position your image elements to the canvas. This remarkable tool utilizes three pretrained models based on advanced versions of Llama3 and Phi3 technologies, which are adept at understanding and translating complex coding into stunning visual content. These models are enriched with diverse datasets, including ground-truth annotations from Open-Images and extracted data from automated image annotations, bolstered by Direct Preference Optimization. This means they not only generate what you code but also learn to understand what makes a visually appealing image based on user feedback. Omost was made by lllyasviel, a legendary creator of IC Light, Fooocus and Controlnet models. For now you can only use it in Gradio but I bet we're gonna see a comfyUI workflow really soon. Github - https://lnkd.in/g2CXR_Yj Also check out a tutorial from Nerdy rodent - https://lnkd.in/gJzYmwuU #omost #LLM #stablediffusion

✨ MicroDreamer: 3D Generation with Zero-Shot Efficiency ✨ 💡 Introduction: Introducing MicroDreamer, a groundbreaking multi-view diffusion model that transforms the landscape of zero-shot 3D generation. This innovative system leverages a score-based iterative reconstruction algorithm to dramatically reduce the number of function evaluations, propelling the production of high-quality meshes within a mere 20 seconds. MicroDreamer stands as a testament to algorithmic ingenuity, delivering a leap in efficiency without relying on additional 3D data and capable of running on a single A100 GPU. ⚙️ Main Features: MicroDreamer's core innovation lies in its Score-based Iterative Reconstruction (SIR) algorithm, which iterates 3D parameters to replicate the reconstruction process and reduce function evaluations. The system also incorporates a hybrid forward process, allowing for optimization within pixel space and refining images rendered from 3D objects. Furthermore, MicroDreamer is compatible with both Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS), supporting text-to-3D and image-to-3D tasks. 📖 Case Study or Example: In practical terms, MicroDreamer can generate a complex 3D model like a NeRF representation of a landscape or a detailed mesh of an artifact in about 20 seconds on standard hardware. This efficiency is game-changing for industries like gaming, where rapid prototyping of 3D models is essential, or in medical imaging, where quick 3D reconstruction can aid in diagnosis and treatment planning. ❤️ Importance and Benefits: MicroDreamer is not just a step but a leap forward in 3D content generation. Its efficiency eclipses traditional optimization-based methods by 5-20 times. This hastens workflows in industries reliant on 3D modeling and democratizes access to high-quality 3D generation for creators with limited resources. The approach's minimal reliance on 3D data also signifies a reduction in the need for extensive and costly datasets. 🚀 Future Directions: While MicroDreamer sets a new standard in efficiency and quality, future improvements could include enhancing consistency models or alternative sampling models to further improve efficiency. Additionally, the model's fidelity and 3D consistency could be improved by advancing the outputs of multi-view diffusion models. 📢 Call to Action: Are you ready to delve into the future of 3D generation? Explore the full potential of MicroDreamer and join the conversation on how this technology can reshape your industry. Access the code and contribute to the evolution of zero-shot 3D generation at [https://lnkd.in/emZyBn5E]. #MicroDreamer #3DGeneration #ZeroShotLearning #DiffusionModels #MachineLearning #AI #Innovation #TechTrends

Data Argumentation techniques for VisionEncoderDecoder ?? -> Understand the probelm from the pre-trained model. is it IMAGE CAPTURING, is it PARSING or it is OCR. If OCR: - Geometric Transformations Rotation, Translation, Scaling, Flipping, Cropping. - Color and Contrast Adjustments - Noise Addition - Blurring - Threshold - Elastic Deformations - Template Matching If Parsing: - Document Layout Variations: Layout Perturbation and Page Orientation - Image-Based Augmentation: Image Cropping and Resizing same as OCR - Multi-Modal Augmentation: Combining Text and Image Augmentation- Combine text perturbation techniques with image-based augmentations to create diverse training samples. and Generate synthetic documents that include varied text content and layouts, combining both text and image augmentation techniques. HOW: OPENCV documentation

📢📢 𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝗛𝘆𝗺𝗯𝗮: 𝗔 𝗡𝗲𝘄 𝗦𝗺𝗮𝗹𝗹 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹 🚀𝗧𝗵𝗲 𝗛𝘆𝗺𝗯𝗮-𝟭.𝟱𝗕 𝗕𝗮𝘀𝗲 𝗺𝗼𝗱𝗲𝗹 𝗻𝗼𝘁𝗮𝗯𝗹𝘆 𝗼𝘂𝘁𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝘀 𝗟𝗟𝗮𝗠𝗔 𝟯.𝟮-𝟯𝗕, 𝗮𝗰𝗵𝗶𝗲𝘃𝗶𝗻𝗴 𝘁𝗵𝗶𝘀 𝘄𝗶𝘁𝗵 𝟳× 𝗳𝗲𝘄𝗲𝗿 𝘁𝗼𝗸𝗲𝗻𝘀 𝗮𝗻𝗱 𝗱𝗲𝗹𝗶𝘃𝗲𝗿𝗶𝗻𝗴 𝗮 𝟭𝟮× 𝗿𝗲𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗶𝗻 𝗰𝗮𝗰𝗵𝗲 𝘂𝘀𝗮𝗴𝗲. 𝗞𝗲𝘆 𝗙𝗲𝗮𝘁𝘂𝗿𝗲𝘀 𝗼𝗳 𝗛𝘆𝗺𝗯𝗮: 🌟𝗛𝘆𝗯𝗿𝗶𝗱 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲: Integrates parallel Mamba and attention mechanisms for optimally balancing performance and efficiency. 🌟𝗠𝗲𝘁𝗮 𝗧𝗼𝗸𝗲𝗻𝘀: Learnable tokens designed to alleviate attention collapse and serve as dynamic knowledge repositories. 🌟𝗦𝗹𝗶𝗱𝗶𝗻𝗴 𝗪𝗶𝗻𝗱𝗼𝘄 𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻: Refines focus while optimizing KV cache handling efficiency. 🌟𝗖𝗿𝗼𝘀𝘀-𝗟𝗮𝘆𝗲𝗿 𝗞𝗩 𝗖𝗮𝗰𝗵𝗲 𝗦𝗵𝗮𝗿𝗶𝗻𝗴: Enhances overall efficiency by reducing redundancy. 𝗠𝗼𝗿𝗲 𝗱𝗲𝘁𝗮𝗶𝗹𝘀 𝗮𝗯𝗼𝘂𝘁 𝗛𝘆𝗺𝗯𝗮:  🔗𝗧𝗲𝗰𝗵𝗻𝗶𝗰𝗮𝗹 𝗥𝗲𝗽𝗼𝗿𝘁: https://lnkd.in/gFKfdG4h 🤗𝗛𝘆𝗺𝗯𝗮-𝟭.𝟱𝗕-𝗜𝗻𝘀𝘁𝗿𝘂𝗰𝘁 𝗠𝗼𝗱𝗲𝗹: https://lnkd.in/e35aatHC 🤗𝗛𝘆𝗺𝗯𝗮-𝟭.𝟱𝗕-𝗕𝗮𝘀𝗲 𝗠𝗼𝗱𝗲𝗹: https://lnkd.in/eRD64N6y