#real-time

GPUPixel offers a versatile, compact library for developers focusing on real-time image and video filtering across multiple platforms. Utilizing C++11 and OpenGL/ES, it includes filters such as skin smoothing and face slimming. Supporting iOS, Android, Mac, Windows, and Linux, GPUPixel enables easy integration and ensures professional quality visuals. Engage with the community for collaborative enhancement and support.

The project enhances the OpenPose algorithm for effective 2D multi-person pose estimation on CPU, maintaining accuracy. It employs a streamlined model to identify 18 keypoints per individual, achieving 40% AP on the COCO 2017 validation set. The repository includes essential code for training, validation, and conversion to OpenVINO format, with C++ and Python demo support, catering to developers seeking efficient pose estimation in limited-resource settings.

This project utilizes a diffusion-based method to generate realistic 3D expressions and gestures in real-time from speech input. It is optimized for Ubuntu and uses dependencies like PyTorch, with models trained on datasets like BEAT and SHOW. Ideal for AI and 3D animation fields, it includes guidance for custom audio input inference and uses tools like Blender for visualization, enhanced by contributions from top researchers.

RobustVideoMatting employs a recurrent neural network to enhance human video matting through temporal memory, facilitating real-time processing without extra inputs. Achieving 4K at 76FPS and HD at 104FPS on an Nvidia GTX 1080 Ti GPU, it supports developers with demonstrations and model downloads for varied frameworks like PyTorch, ONNX, and TensorFlow for seamless integration.

This project delivers a PyTorch and Taichi-driven implementation of the instant-ngp NeRF training pipeline. It includes guides for installation, dataset processing, and training scripts applicable to both preprocessed datasets and custom videos. Compatibility with Linux and RTX graphics cards is emphasized, with half2 optimization available for improved performance. Mobile deployment is feasible through Taichi AOT, enabling real-time iOS rendering. The project also supports text-to-3D rendering as a backend for the stable-dreamfusion project. Explore streamlined procedures and comprehensive instructions for efficient NeRF training and deployment.

AudioDec is an open-source audio codec delivering high-quality 48 kHz mono speech at 12.8 kbps. It efficiently balances low latency and high compression, suitable for real-time streaming on both GPU and CPU. The codec adapts quickly to new applications through a two-stage training process, making it ideal for telecommunication with excellent audio reconstruction and minimal delay.

Darkflow utilizes the YOLO framework to deliver efficient real-time object detection and classification. Compatible with Python 3, TensorFlow, and OpenCV, it provides GPU and CPU support. Features include easy installation, flexible model configurations, and JSON output, making it ideal for scalable object detection across applications.

Explore a pitch estimation method designed with self-supervised transposition-equivariant objectives, noted for its high accuracy and fast processing. Built with PyTorch, PESTO provides user-friendly integration through both a command-line interface and a Python API, and supports batch processing across diverse audio formats. Tailored for swift pitch analysis, PESTO offers multiple export options and functions effectively even without advanced hardware, making it applicable for both research-oriented and practical music processing scenarios.

The project uses an advanced framework to enhance real-time object detection across different tasks. By integrating models like YOLOR-CSP, YOLOR-CSP-X, and YOLOR-P6, the project shows significant improvements in Average Precision metrics on COCO datasets. It employs innovative features, offering enhanced processing speed and accuracy, making it a valuable tool for researchers and developers.

NanoDet-Plus is a lightweight object detection model known for its speed and accuracy, especially on mobile platforms. It supports backends such as ncnn, MNN, and OpenVINO and offers up to 34.3 mAP and 97fps on mobile ARM CPUs. With its Assign Guidance Module and Dynamic Soft Label Assigner, the model significantly improves accuracy without extensively using GPU resources. These attributes make it a suitable choice for various object detection needs in real-time applications.

YOLOv10 advances real-time object detection by improving architectural efficiency and eliminating NMS, offering a balanced design for speed and accuracy. This PyTorch implementation notably outperforms previous models, enhancing performance while reducing computational demands, ideal for applications demanding swift, efficient detection.