Coding & Development

AI Dev Gallery is an open-source project from Microsoft designed for Windows developers to integrate AI capabilities into their applications. It provides a comprehensive learning resource with over 25 interactive samples powered by local AI models. Developers can easily browse, download, and run various AI models directly from platforms like Hugging Face and GitHub. The gallery also allows users to view the C# source code for samples and export standalone Visual Studio projects with a single click, facilitating hands-on learning and integration. It supports offline use once models are downloaded and features popular open-source models and APIs from the Microsoft Foundry on Windows. The project is completely open-source, encouraging contributions and feedback from the developer community.

E2E FT Marigold for Normals is an AI tool hosted on Hugging Face that specializes in generating surface normals from uploaded images. Users can input an image and receive two outputs: the raw data of the surface normals and a corresponding colored map. This tool is particularly useful for tasks requiring detailed surface information, such as 3D reconstruction, computer vision research, or graphics applications. It is licensed under Apache-2.0, making it accessible for various projects. The platform leverages Hugging Face's infrastructure, which offers different pricing tiers for storage, compute, and inference, catering to both individual developers and enterprise teams.

COCO-WholeBody is a comprehensive dataset designed for whole-body human pose estimation, building upon the COCO 2017 dataset. It offers extensive annotations for 133 keypoints per person, covering 17 for the body, 6 for feet, 68 for the face, and 42 for hands, along with bounding boxes for the person, face, and each hand. This dataset is crucial for researchers and developers working on advanced computer vision tasks, particularly in human pose analysis. The project provides evaluation tools and has been utilized in top-tier computer vision conferences, making it a valuable resource for academic and non-commercial research in the field.

pytorch-openpose offers a PyTorch implementation of the popular OpenPose framework, enabling robust body and hand pose estimation. This tool is particularly useful for researchers and developers working with computer vision and human-computer interaction, as it facilitates the conversion of existing OpenPose Caffe models to PyTorch. Beyond body and hand pose, it also supports the implementation of face keypoint detection, following similar procedures used for hand detection. The project provides clear instructions for setting up a Python 3.7 environment, installing necessary dependencies, and downloading pre-trained models. Demos are available for webcam, image, and video processing, making it accessible for immediate experimentation and integration into various projects.

pytorch-yolo-v3 offers a PyTorch implementation of the YOLO v3 object detection algorithm, designed for efficient and real-time object recognition. This repository aims to improve upon existing ports by streamlining the code, removing redundant components, and providing clear documentation. It currently supports detection in single images, multiple images, and video streams, with options to adjust resolution and utilize half-precision floats for faster inference. The project serves as a driver code for research, with plans to include a training module in the future. It requires Python 3.5, OpenCV, and PyTorch 0.4.

Pipeless is an open-source computer vision framework designed to accelerate the development and deployment of AI applications. It abstracts away complexities like code parallelization, multimedia pipelines, memory management, and model inference, allowing developers to build and deploy real-time computer vision applications rapidly. Inspired by serverless technologies, Pipeless enables users to define 'stages'—micro-pipelines that perform specific tasks. These stages can be dynamically combined per stream, supporting multi-stream processing and on-the-fly configuration changes. It supports industry-standard models and custom models across various inference runtimes like ONNX Runtime, CUDA, TensorRT, and OpenVINO, ensuring high performance on both CPU and GPU. Pipeless also offers multi-language support for hooks and built-in restart policies for robust operation on edge, IoT, or cloud environments.

PoseEstimationForMobile is an open-source project designed for real-time single-person pose estimation on Android and iOS devices. It leverages CPM and Hourglass models, implemented with TensorFlow, and incorporates inverted residuals (MobileNet V2) for optimized, real-time inference. The repository includes code for training both CPM and Hourglass models, along with demo source code for Android and iOS. This allows developers to integrate pose estimation capabilities into their mobile applications with high performance. The project provides pre-trained models and detailed instructions for setting up training environments, converting models for mobile deployment (Mace, TFLite, CoreML), and benchmarking performance across various mobile chipsets.

PVN3D is the official source code for "PVN3D: A Deep Point-wise 3D Keypoints Hough Voting Network for 6DoF Pose Estimation," a research paper presented at CVPR 2020. This open-source project enables researchers and developers to implement and experiment with advanced 6DoF pose estimation techniques using 3D keypoints. It supports training and evaluation on popular datasets like LineMOD and YCB-Video, and includes pre-trained models for various objects. The tool also offers guidance for adapting the framework to new datasets, making it a valuable resource for academic research and development in computer vision and robotics. It is built with Python and PyTorch, requiring specific CUDA and Python environment setups.

Open Free - AI Playground is a curated collection of free AI services, presented as a Hugging Face Space. Users can browse various AI playgrounds by categories such as Popular, Best, or New. Each AI demo is showcased directly within an embedded frame, allowing for live interaction, or as a screenshot when live embedding is not feasible. This platform serves as a central hub for exploring and interacting with a diverse range of AI tools and models, making it accessible for anyone interested in experiencing different AI capabilities without cost. It's an ideal resource for quick experimentation and discovery of new AI applications.

ReinforcementLearning.jl is a comprehensive open-source package designed for reinforcement learning research within the Julia programming language. It emphasizes reusability and extensibility, offering elaborately designed components and interfaces that simplify the implementation of new algorithms. The package also facilitates easy experimentation, allowing users to run benchmark experiments, compare different algorithms, and evaluate agents efficiently. A core focus is on reproducibility, supporting a range of methods from traditional tabular approaches to modern deep reinforcement learning algorithms. It integrates several sub-packages like ReinforcementLearningBase.jl, ReinforcementLearningEnvironments.jl, and ReinforcementLearningCore.jl to provide a robust and modular framework for researchers and developers.

Spinning Up in Deep RL is an educational resource developed by OpenAI designed to simplify the learning process for deep reinforcement learning (deep RL). This comprehensive module offers a short introduction to RL terminology, various types of algorithms, and fundamental theory. It also includes an essay on how to transition into an RL research role, a carefully curated list of important research papers organized by topic, and a well-documented code repository featuring concise, standalone implementations of key algorithms. Additionally, it provides several exercises to serve as warm-ups, making it an ideal starting point for individuals looking to understand and apply deep reinforcement learning concepts. The resource is currently in maintenance mode, focusing on bug fixes and minor updates.

second.pytorch is an open-source project providing a SECOND detector for object detection, specifically designed for KITTI and NuScenes datasets. It leverages sparse convolution-based networks for efficient processing. The tool supports Python 3.6+ and PyTorch 1.0.0+, and has been tested on Ubuntu 16.04/18.04 and Windows 10. Key features include support for NuScenes, PointPillars, fp16 mixed precision, and multi-GPU training. The project also offers a KITTI viewer for data visualization and evaluation. While the project is currently deprecated in favor of OpenPCDet or mmdetection3d, it remains a valuable resource for understanding and implementing SECOND-based object detection.

Bioclip Demo is an interactive application hosted on Hugging Face Spaces, designed for running BioCLIP inference on images of living organisms. Users can upload a picture and either select a taxonomic level (e.g., genus, species) or provide custom class names. The tool then returns the most likely names along with confidence scores, making it valuable for visualization, data exploration, and biological research. It supports tasks such as zero-shot image classification, aiding in the identification and categorization of species based on visual input. This demo is part of the HDR Imageomics Institute's efforts to make advanced AI models accessible for scientific applications.

LHM, or Large Animatable Human Model, is an open-source tool available as a Hugging Face Space that enables users to generate and animate 3D human avatars. Users can upload a full-body or half-body photo of a person and then select a driving video to provide the desired motion. The application constructs a 3D human avatar from the input image and subsequently animates it according to the movements in the chosen video. This tool is ideal for those looking to create custom animated human models for various applications, from research to creative projects, leveraging its capabilities for 3D reconstruction and animation.

Latent Consistency Models (LCMs) is an open-source platform hosted on Hugging Face Spaces, designed for AI researchers and developers to explore and experiment with latent consistency models. While the live application currently displays a runtime error, the underlying project aims to provide a space for engaging with advanced AI algorithms. This tool is particularly relevant for those interested in the technical aspects of AI model development and research, offering a foundation for understanding and manipulating latent spaces within AI models. It serves as a community-driven resource for advancing knowledge in the field of AI.

MEGA-Bench Leaderboard is a comprehensive platform designed for evaluating multimodal AI models. Hosted on Hugging Face, this tool provides users with detailed performance metrics and allows for easy comparison of various models. Users can select different tables and apply filters to view specific data, making it an invaluable resource for researchers and developers in the AI community. The platform aims to offer transparency and a standardized way to benchmark the capabilities of multimodal models, contributing to advancements in the field. It is freely accessible, promoting open research and collaboration.

Making Demos Leaderboard is a Hugging Face Space designed to track and showcase AI demos. It provides a dynamic leaderboard that ranks submissions based on the number of likes they receive from the community. This platform encourages participation in the 'Making Demos' event and allows users to see top-performing AI demonstrations. While currently paused, the tool aims to foster community engagement and provide a competitive yet collaborative environment for AI enthusiasts to share and discover innovative projects. Users can typically refresh the leaderboard to view updated rankings and explore various AI applications.

MotionBench Leaderboard is an open-source platform designed for the evaluation and comparison of various motion models. Users can submit their model evaluation JSON files to the leaderboard, which then allows for comprehensive analysis and benchmarking. The platform provides functionalities to view and filter the leaderboard data based on different evaluation dimensions, making it easy to track progress and identify top-performing models. Additionally, users have the convenience of downloading the entire leaderboard as a CSV file for further offline analysis or integration into other systems. This tool is ideal for researchers and developers in the AI community who need a standardized way to assess and compare the performance of their motion-related AI systems.

The MTEB Legacy Leaderboard offers a comprehensive platform for evaluating and comparing text embedding models. Users can access an archived leaderboard to search for specific models, filter results by model type or size, and view sortable tables displaying each model's scores across various benchmarks. This tool is designed to help AI researchers and developers assess the performance of different AI systems in understanding and representing text, providing valuable insights into model capabilities and tracking progress within the AI community. It serves as a crucial resource for benchmarking and understanding the landscape of text embedding models.

The Multimodal Hallucination Leaderboard is a Hugging Face Space developed by Typhoon AI, designed for evaluating and comparing the hallucination tendencies of various multimodal AI models. Users can access and explore existing results from established AI hallucination benchmarks such, as POPE/MHaluBench and AVHalluBench. The platform also provides functionality for users to submit their own evaluation results, contributing to a broader understanding of AI model performance. This tool is particularly valuable for researchers and developers focused on understanding, benchmarking, and ultimately mitigating inaccuracies and hallucinations in AI outputs across different modalities.

The MMLU-Pro Leaderboard, hosted on Hugging Face Spaces by TIGER-Lab, provides a platform for evaluating and comparing the performance of AI models on more advanced and challenging multi-task evaluations. Users can easily search and filter model data based on various criteria such as model name, parameter size, and specific subjects. The tool also offers customization options for displayed columns, allowing researchers and developers to tailor the view to their specific needs. This leaderboard is designed to offer insights into model capabilities on complex tasks, making it a valuable resource for academic research and AI development.

Model Memory Utility is a practical AI tool designed to assist developers and engineers in managing and optimizing the memory usage of AI models. This application, hosted on Hugging Face Spaces, allows users to estimate the video memory required for both training and inference with models sourced from the Hugging Face Hub. By simply entering the model name or URL, selecting the relevant library, and specifying desired precisions (e.g., float16, float32), users can gain crucial insights into memory requirements. This capability is essential for tuning model performance, optimizing resource allocation, and facilitating efficient cloud deployment, ultimately helping to prevent out-of-memory errors and reduce operational costs.

On Device Demo is a demonstration tool built on Hugging Face Spaces, showcasing the capabilities of running AI models directly on a user's device. Utilizing the Ratchet and Whisper frameworks, this tool enables local execution of models, which results in faster processing and improved efficiency compared to cloud-based solutions. It functions as a toolkit for developers and researchers interested in on-device AI, eliminating the need for specific input beyond the initial setup. The demo highlights the potential for enhanced privacy and reduced latency by keeping computations local. It's an excellent resource for understanding the practical application of Ratchet Whisper in a real-world scenario.

The Open VLM Leaderboard, hosted on Hugging Face, provides a comprehensive platform for viewing and analyzing the performance of various vision-language models (VLMs). It aggregates evaluation results from the VLMEvalKit benchmark, offering a centralized resource for researchers and developers. Users can easily narrow down results by selecting specific evaluation dimensions, filtering by model size or type, or searching for a particular model name. This tool is designed to facilitate the comparison and understanding of VLM capabilities, aiding in the development and selection of appropriate models for different applications. It serves as a valuable resource for anyone working with or interested in the advancements of vision-language AI.