Google DeepMind’s Post

TIL23 OpenAI's Sora, text-to-video model, amazing demos possibilities. Generate videos up to 1min at 1080p, various aspect ratios and resolutions. Still struggles with accurately simulating physics of a complex scene. Trained using synthetic captions that massively optimized the training process. Scaling up # of images / frames of images from training videos -> better results. https://lnkd.in/e5FYkcrX

For video stabilization and motion tracking, SIFT has some impressive results! I used SIFT in OpenCV and feature matching in this demo. I also found a paper that talked about this method. https://lnkd.in/e96iSSpW check out this demo that I wrote here https://lnkd.in/ekBF4pFz

Excited to Share: Helmet Detection Project Using OpenCV and YOLO! Key Components:: Object Detection with YOLO: We implemented the YOLO algorithm, known for its real-time processing capabilities, to accurately detect and locate helmets in images and videos. OpenCV Integration: OpenCV, a versatile computer vision library, played a pivotal role in preprocessing, image manipulation, and post-processing steps. Its flexibility allowed us to optimize and fine-tune the detection pipeline. Challenges Overcome: Dealing with variations in lighting and environmental conditions. Addressing challenges related to different helmet types and orientations. #AIMERS #MACHINELEARNING #AIMERSOCIETY #YOLO #OPENCV

FACIAL EMOTION DETECTION step-by-step tutorial - Part 3. In this part, I apply the object tracking algorithm SORT over the existing model to track the movement of the facial regions. The accuracy on the private_test_set is also improved by 1% (compared to the last video) using some parameter initialization technique and a new learning rate scheduler.

OpenAI just announced Strawberry - its latest model. Here are 5 insane demos from the event: 1. You can simply code a video game from one prompt - https://lnkd.in/ei8G4rV8 2. You can now solve complex logic puzzles easily - https://lnkd.in/eJ2FK9aK 3. Easily understand & translate corrupted sentences - https://lnkd.in/ehvHBPzf 4. Understand more complex math equations - https://lnkd.in/evGpw42f 5. Skynet? 😵 Oh, it also knows how many "R"s there are in "Strawberry."

Unlock the world of geometry with the MC4.0 AIOT Kit! Watch how this amazing tool makes learning shapes, angles, and spatial reasoning fun and interactive. Perfect for students and educators looking to bring math to life. Let's make geometry exciting! 🚀🔢 #GeometryLessons #STEAMKit #MathEducation #InteractiveLearning #EdTech #HandsOnMath #MC4.0 #ShapeUpYourSkills #LearningGeometry #FutureMathematicians"

Excited to share a fascinating project I've been working on: real-time hand tracking using OpenCV and MediaPipe! 🖐️ This project showcases the power of computer vision by detecting and visualizing hand landmarks in a live video feed. It's an essential tool for various applications, including gesture recognition, interactive interfaces, and beyond. key features: ✅Real-Time Processing ✅Landmark Detection ✅OpenCV and MediaPipe Integration

New Open Vocabulary YOLO-World model has been released🚀!! 💯 YOLO-World is a zero-shot object detector that can be used for various situations and different features along images. 🔥 It is very fast in compare with other zero-shot detectors like GroundingDINO. 📖 Custom classes can be defined by the user and then exported. In this way the model will keep the selected classes for next time to predict. Thanks to Ultralytics framework we can easily run YOLO-World without any difficult installation process. #deeplearning #deeplearningai #machinelearning #yolov8 #yolo #opencv #computervision

Why do we needs 100-1000s of demos to train even simple robot tasks? The answer: Supervised Learning wastes rich observational information. To fix this, we built DynaMo, a Self-Supervised method that operates on small in-domain data by exploiting the dynamics of temporal data. Instead of learning a denoising or contrastive objective on a bag of frames, we exploit the sequential and causal structure of the demonstration observations. DynaMo jointly learns inverse dynamics to a latent action, and forward dynamics predicting the next frame embedding. It requires no augmentations, contrastive sampling, or ground truth actions. We are releasing our open-source code, data, and videos here: https://lnkd.in/e5TFR49w . Try it out! This work is led by Zichen Jeff Cui w/ Hengkai Pan, Aadhithya Iyer and Siddhant Haldar.

My first experience on a Kaggle competition solving a classification problem using ScikitLearn and analysis tools.