Duc Liem Dinh’s Post

🤔Can #AI help people to sleep better? BodyMap the new #method developed by Carnegie Mellon University can predict the #3D posture of the human body and its pressure in bed. BodyMap works in three steps: 1️⃣It analyzes a depth image of the body in bed and its pressure image from a special mattress 2️⃣It creates a detailed 3D model of the body, with precise shape, posture and pressure on each point 3️⃣It aligns pressure images with 3D model using AI, improving reconstruction and localization of high-pressure areas BodyMAP exceeds the #state of the art by 25% in terms of accuracy for both body mesh prediction and applied 3D pressure map, using real data on people in bed. 👉Follow us if you want to know what's new in the world of AI. NVIDIA, Abhishek Tandon, Anujraaj Goyal, Henry Clever, Zackory Erickson Paper 👉🏻 https://lnkd.in/dRrbnpYm Project 👉🏻 https://lnkd.in/d5mEa9kE

Try it out: It turns out AI is bad at creating plain white paintings In the dynamic interplay of technology and creativity, a recent exploration reveals a captivating anomaly: AI's struggle with the minimalist concept of creating a plain white painting. This insight, derived from research conducted by data scientist Cody Nash, underlines a fascinating frontier in AI's evolution. Despite monumental advancements in AI art generation, exemplified by technologies like Sora AI and Nvidia's Blackwell AI superchip, AI demonstrates a notable challenge in mastering the simplicity of minimalist art. Nash's experiments with DALLE 3 and Stable Diffusion XL, aimed at generating an all-white image, consistently resulted in outputs that, while technically meeting the criteria, lacked the depth and emotional resonance inherent in human-created minimalist pieces. This intriguing limitation not only highlights the distinct gap between AI-generated and human-created art but also reassures the unique value of human creativity in the realm of art. As CAIOs, this narrative serves as a poignant reminder of the nuanced complexities of integrating AI within creative domains, urging a balanced approach that values both technological innovation and the irreplaceable essence of human creativity. #GenAI #Creativity #Leadership #Innovation #DigitalTransformation #AIChallenges #HumanCreativity #ArtAndTechnology https://lnkd.in/d4qz4Swb

Advancing the field of human mesh recovery, the introduction of GenHMR marks a significant milestone in computer vision applications, particularly in health, arts, and entertainment. Despite historical challenges in extracting 3D human poses from monocular images due to depth ambiguity and occlusions, traditional methods and probabilistic approaches have faced limitations. GenHMR revolutionizes monocular human mesh recovery by framing it as an image-conditioned generative task. This innovative framework addresses uncertainties in 2D-to-3D mapping through two key components: 1. Pose Tokenizer: Converting 3D poses into discrete tokens within a latent space. 2. Image-Conditional Masked Transformer: Learning probabilistic pose token distributions based on image cues and masked token sequences. By decoding high-confidence pose tokens iteratively and employing a 2D pose-guided refinement approach, GenHMR ensures accurate alignment of 3D body meshes with 2D pose references. The groundbreaking results speak for themselves - GenHMR surpasses existing methodologies on standard datasets, establishing a new benchmark in 3D reconstruction precision. For further insights, explore Muhammad Usama Saleem's research on GenHMR at: https://lnkd.in/dF4Tq3yY Join me as I delve into the future prospects of generative AI. Your valuable thoughts on this advancement are highly appreciated. #AI #ComputerVision #HumanMeshRecovery #GenerativeAI #ResearchInnovation

🪬🪬UHM: Authentic Hand Avatar by Phone🪬🪬 👉 META unveils UHM, novel 3D high-fidelity avatarization of your (yes, the your one) hand. Adaptation pipeline fits the pre-trained UHM via phone scan. Source Code released under CC NonCommercial 4.0💙 𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬: ✅UHM a novel neural 3D hand model ✅Universal HD 3D hand meshes of arbitrary IDs ✅Adapting to each person by quick phone scan ✅UHM performs tracking & modeling at same time ✅Novel image matching loss function for skin sliding issue #artificialintelligence #machinelearning #ml #AI #deeplearning #computervision #AIwithPapers #metaverse 👉Discussion https://lnkd.in/dMgakzWm 👉Paper https://lnkd.in/dyGaiAnq 👉Code https://lnkd.in/d9B_XFAA

The prompt used for generating the latest image was: "Illustrate a close-up of a robot's hand shaking hands with a human, emphasizing the collaborative relationship between humans and generative AIs. The scene merges futuristic elements with the style of ancient cave paintings, highlighting control and partnership. The background should feature symbols representing various industries influenced by AI, like airplanes for aeronautics, stethoscopes for medicine, gavels for legal, books for education, banks for banking, and cars for the vehicle industry, all interconnected by a central AI symbol. This symbol and the handshake should depict the AI as a harmonious tool within human grasp, symbolizing control and collaboration. Replace the title 'AI: The Future Unfolded' with 'The Future is Here and Now', integrating it into the scene in an ancient script aesthetic that is clear and legible, conveying the message of harmonious advancement." #ChallengeAccepted #CreateAPosterForAMovieAboutTheFutureOfAI #LearnTechWithLinkedIn #PromptEngineering #AI Ronnie Sheer

𝙁𝙞𝙣𝙙𝙞𝙣𝙜 𝙀𝙙𝙜𝙚𝙨 𝙇𝙞𝙠𝙚 𝙖 𝙋𝙧𝙤: 𝙏𝙝𝙚 𝙈𝙖𝙜𝙞𝙘 𝙤𝙛 𝘼𝙘𝙩𝙞𝙫𝙚 𝘾𝙤𝙣𝙩𝙤𝙪𝙧 𝙈𝙤𝙙𝙚𝙡𝙨 ✨ Ever seen an algorithm that chases shapes around like it’s their best friend? 👀   Yep, that’s an 𝗔𝗰𝘁𝗶𝘃𝗲 𝗖𝗼𝗻𝘁𝗼𝘂𝗿 𝗠𝗼𝗱𝗲𝗹 for you! 💡 𝗪𝗵𝗮𝘁’𝘀 𝘁𝗵𝗲 𝗳𝘂𝘀𝘀 𝗮𝗯𝗼𝘂𝘁?   Think of this as a “shape-hugging” algorithm. It’s like the algorithm and the object shape are dancing together, getting closer and more accurate with each step. 💃🕺 Why should we care? Well, these contour models have a special superpower—they’re great at finding boundaries in images, which is super helpful for everything from medical imaging to object tracking in videos. 🎯 🔍 𝗛𝗼𝘄 𝗱𝗼 𝘁𝗵𝗲𝘆 𝘄𝗼𝗿𝗸?   Active Contour Models (or Snakes, as they are sometimes called) work by evolving a curve towards the edges of objects. It’s like drawing a line around an object, but with math and a little bit of magic! ✨ 📌 𝗛𝗲𝗿𝗲’𝘀 𝗵𝗼𝘄 𝘁𝗵𝗲𝘆 𝗯𝗿𝗲𝗮𝗸 𝗶𝘁 𝗱𝗼𝘄𝗻:   1️⃣ 𝗜𝗻𝗶𝘁𝗶𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻   The model starts with an initial contour or boundary. It’s a rough guess of where the object might be. Like drawing with your non-dominant hand 🤚 (not very precise at first).  2️⃣ 𝗘𝘅𝘁𝗲𝗿𝗻𝗮𝗹 𝗙𝗼𝗿𝗰𝗲𝘀   External energy pulls the contour toward the object’s edges. This energy comes from image gradients, so the stronger the contrast, the faster the contour moves. Imagine the contour being dragged toward the object’s edges by invisible hands 👐. 3️⃣ 𝗜𝗻𝘁𝗲𝗿𝗻𝗮𝗹 𝗙𝗼𝗿𝗰𝗲𝘀   While external forces try to match the contour to the object, internal forces try to keep the shape smooth, preventing it from becoming too wild or jagged. It’s a balancing act 🤹♂️ between accuracy and simplicity! 4️⃣ 𝗧𝗵𝗲 𝗙𝗶𝗻𝗮𝗹 𝗦𝗵𝗮𝗽𝗲   After several iterations, the contour aligns with the object’s boundary, like finding the perfect fit for a glove 🧤. The shape is now locked in. 🔥 𝗪𝗵𝗲𝗿𝗲 𝗰𝗮𝗻 𝘆𝗼𝘂 𝘂𝘀𝗲 𝗶𝘁?   - Medical Imaging 🏥: Detecting tumors, vessels, or organs in scans.   - Video Tracking 🎥: Following moving objects in video frames.   - Object Recognition 🖼️: Drawing boundaries for identifying shapes in images. #ComputerVision #ActiveContours #MachineLearning #AI #ImageProcessing #Shapeshifter #Contours #MedicalImaging #Innovation

Why do generative AI models struggle to create realistic videos with hands and faces? The complexity first arises from the anatomical richness of hands and faces. These are not simply static surfaces, but dynamic systems of muscles, tissues, and joints that move with incredible fluidity and variability. Generative AI models essentially operate through a statistical prediction process. They can recognize patterns, but struggle to reconstruct the true dynamic essence of a gesture. The problem amplifies when moving from individual frames to video sequences. AI must not only generate a realistic image but ensure that each successive image is coherent with the previous one, maintaining a movement logic that is natural and fluid. Hands, in particular, represent a true computational nightmare. A single error can immediately transform the image into something grotesque and unnatural. We will need to create AI systems that not only see images but understand them at an almost biomechanical level, simulating the complexity of human movements. #artificialintelligence #video #images

🔫🔫 Free-Moving Object Reconstruction 🔫🔫 👉EPFL (+#MagicLeap) unveils a novel approach for reconstructing free-moving object from monocular RGB clip. Free interaction with objects in front of a moving cam without relying on any prior, and optimizes the sequence globally without any segments. EDIT: code to be released soon (author announced in the comments below). 𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬: ✅Joint reconstruction/pose of free-moving objects ✅Globally optimized with a single network only ✅OTS-2D segmentation for masks in each frames ✅Novel simple-but-effective strategy with virtual cam #artificialintelligence #machinelearning #ml #AI #deeplearning #computervision #AIwithPapers #metaverse 👉Discussion https://lnkd.in/dMgakzWm 👉Paper https://lnkd.in/dYRmv9sP 👉Project https://lnkd.in/dmaiBnn3 👉Code https://lnkd.in/dJRbCu6M

🚀 Excited to share my new preprint: FaceGemma: Enhancing Image Captioning with Facial Attributes for Portrait Images 🎉 We introduce FaceGemma, a model designed to capture and describe nuanced facial attributes for more descriptive, accessible image captions. Using FaceAttdb data and fine-tuning PaliGemma with Llama 3 - 70B, we generated captions for faces, showcasing details like facial attributes (in black), face wears i.e.; eyeglasses (in green), and facial expressions (in red). Priprint: https://lnkd.in/gWS9z4X4 Results: FaceGemma achieved a BLEU-1 score of 0.364 and a METEOR score of 0.355, demonstrating the impact of facial attributes on enhanced captioning accuracy. 🌟 #AI #ComputerVision #ImageCaptioning #LLM

I shared an image today from a source I trust without doing a background check. This picture is widely spreading. This image is #AI-generated due to the following key factors: - Unnatural Soldier Alignment: The soldiers are positioned too uniformly, with little variation in posture or movement, which is unlikely in a real-life raid scenario. - Identical Gear: The soldiers' equipment (helmets, backpacks, and knee pads) looks almost identical, which #AI often struggles with when trying to add natural variation to objects. The shoe color is different from the ones the soldiers are wearing in the videos Al Jazeera shared. - Flat Lighting: The lighting lacks depth, with minimal shadow variation. - Distorted Text: The "Al Jazeera" text on the doors shows slight distortions, which is common in AI-generated #images. - Awkward Weapon Handling: The way some soldiers hold their weapons seems weird, as AI struggles with fine details like hand-object interactions. When it comes to AI-generated images, body parts are often a sign of manipulation. AI can struggle with rendering complex details like hands, fingers, and facial features, which may appear distorted or unnatural. For instance, hands may have the wrong number of fingers. These inconsistencies in body parts are one of the clearest ways to spot AI-generated images, as the #technology often fails to replicate the subtleties and variations found in real human anatomy.