Research & Education

GeoSeg is an open-source semantic segmentation toolbox built on PyTorch, PyTorch Lightning, and timm, primarily focused on developing advanced Vision Transformers for remote sensing image segmentation. It provides a unified benchmark and training script for various segmentation methods, making it simple and effective for further development. The tool supports key remote sensing datasets such as ISPRS Vaihingen and Potsdam, UAVid, LoveDA, and OpenEarthMap. GeoSeg also includes support for multi-scale training and testing, and inference on huge remote sensing images. It integrates a variety of networks including Mamba, PyramidMamba, UNetFormer, DC-Swin, BANet, ABCNet, A2FPN, and CNNs, offering a comprehensive solution for satellite, aerial, and UAV image segmentation.

Intro2Petro serves as an interactive guide for students interested in petroleum engineering, helping them explore various domains within the field based on their interests. The platform offers a structured approach to understanding key areas such as Reservoir Engineering, Drilling Engineering, Production Engineering, Geology, Geophysics, Well Testing, Formation Evaluation, Offshore Engineering, Geomechanics, and Enhanced Oil Recovery. Additionally, Intro2Petro curates a list of essential books, making it a valuable resource for students seeking to deepen their knowledge and identify relevant study materials in petroleum engineering.

indrnn provides a TensorFlow implementation of Independently Recurrent Neural Networks (IndRNN), based on the paper 'Building A Longer and Deeper RNN' by Shuai Li et al. This implementation allows for the creation of longer and deeper recurrent neural networks by ensuring neurons in recurrent layers are independent. A key feature is the element-wise vector multiplication for recurrent weights, where each neuron has a single recurrent weight connected to its last hidden state. This design effectively prevents vanishing and exploding gradients, especially when used with ReLU activation functions, and facilitates stacking multiple recurrent layers. The tool includes examples for reproducing experiments like the Addition Problem and Sequential MNIST.

MaskDINO is an official implementation of the paper "Mask DINO: Towards A Unified Transformer-based Framework for Object Detection and Segmentation," accepted at CVPR 2023. This open-source project offers a unified architecture capable of performing object detection, panoptic segmentation, instance segmentation, and semantic segmentation. It supports task and data cooperation between detection and segmentation, delivering state-of-the-art performance on major datasets like COCO, ADE20K, and Cityscapes. The framework is built upon detectron2 and offers a detrex version. Key features include a flexible architecture where users can easily replace backbone, pixel decoder, and transformer decoder components, and it supports mask-enhanced box initialization for improved performance.

RefineDet is an open-source implementation of a single-shot refinement neural network designed for object detection tasks. Published at CVPR 2018, this method aims to surpass the accuracy of traditional two-stage object detection approaches while preserving the computational efficiency characteristic of one-stage methods. The repository provides comprehensive code for training and evaluating RefineDet models on various datasets, including PASCAL VOC and MS COCO. Users can leverage pre-trained models based on VGG-16 and ResNet-101 architectures, and the system supports both single-scale and multi-scale testing strategies. It includes detailed instructions for installation, data preparation, training, and evaluation, making it a valuable resource for researchers and developers in computer vision.

SNIPER, also known as AutoFocus, is an efficient multi-scale training and inference algorithm designed for instance-level recognition tasks such as object detection and instance-level segmentation. It significantly speeds up multi-scale training by selectively processing context regions around ground-truth objects, called 'chips', operating on low-resolution data. This memory-efficient design allows SNIPER to benefit from Batch Normalization and larger batch-sizes on a single GPU. AutoFocus, the inference component, employs a coarse-to-fine approach, processing only regions likely to contain small objects at finer scales using 'FocusPixels' to generate compact 'FocusChips'. The tool supports half-precision training with no loss in accuracy and offers fast inference speeds, making it suitable for advanced computer vision research and development.

EarlyDiagnostics (EarlyDx) is dedicated to advancing cancer detection and precision medicine through its innovative liquid biopsy tests. The company has developed CancerRadar, a highly accurate and sensitive blood test for detecting and locating cancer. Additionally, EarlyDx offers EarlyDx-Cloud, a comprehensive bioinformatics solution designed to assist diagnostic companies in offering liquid biopsy tests for cancer detection, therapy selection, and treatment monitoring. Their core technologies include cfMethyl-seq for profiling cfDNA methylome, cfSNV for accurately calling mutations from blood, and cfTrack for non-invasive monitoring of cancer MRD, recurrence, and evolution. EarlyDx aims to provide accessible and reliable tools for early disease diagnostics.

vecstack is a Python package designed for implementing stacking, a powerful machine learning ensembling technique also known as stacked generalization. It provides a convenient way to automate out-of-fold (OOF) computation, prediction, and bagging across various models. The package features a minimalistic functional API for quick integration and a standardized scikit-learn compatible API, allowing for seamless use within existing scikit-learn pipelines, including multilevel stacking with `sklearn.pipeline.Pipeline` and `FeatureUnion`. It supports classification and regression tasks, handles class labels or probabilities, and allows for user-defined metrics and transformations. vecstack is RAM-friendly and can automatically save stacked features and hyperparameters, making it suitable for competitive machine learning environments like Kaggle.

Volatile AI specializes in molecule and volatile organic compound analytics, offering field instrumentation and software for end-users. The company focuses on volatilomics fingerprinting, enabling rapid chemical analysis without complex sample preparation. Their technology includes accessible gas chromatography and custom portable instruments designed for various applications such as biopharmaceutical fermentation monitoring, asphalt variant profiling, and whiskey maturation monitoring. Volatile AI positions itself as a leading research company in this field, providing solutions for understanding molecular composition outside of traditional lab settings.

Mach42 is an AI-driven acceleration company focused on developing ultra-fast AI tools to address complex challenges in computation, particularly within the semiconductor industry. Its Discovery Platform offers solutions for silicon verification, enabling design exploration in days rather than months through accelerated search. The platform also facilitates automatic generation of surrogate models compatible with SPICE simulators for circuit model creation. Additionally, Mach42 provides AI-driven simulation tools for multi-physics challenges. The company aims to deliver higher quality predictions at orders-of-magnitude faster speeds, leveraging AI to accelerate expensive calculations with minimal data and high accuracy.

NyBerMan Bioinformatics Europe is a leading platform dedicated to innovative bioinformatics learning and solutions. It offers a comprehensive suite of services including hybrid learning programs, workshops, internships, and certifications in bioinformatics. The platform covers various aspects such as sequence analysis (NGS, multi-omics integration, genome assembly), structural bioinformatics (AI/ML aided drug discovery, molecular modeling, virtual screening), and functional genomics analysis. Additionally, NyBerMan provides bioinformatics services enhanced by state-of-the-art data processing and visualizations, and develops AI-driven computational platforms to address complex bioinformatics challenges.

The Computer Vision Lab at ETH Zürich is a leading research and education institution dedicated to advancing the field of computer vision. The lab focuses on the computer-based interpretation of both 2D and 3D image data, addressing complex challenges in various domains. Key research areas include medical image analysis, object recognition, gesture analysis, and comprehensive scene understanding. The lab's mission involves developing universal concepts and methods that can be applied across diverse applications. It actively collaborates with other academic labs and industry partners to translate its cutting-edge research into practical solutions and innovations.

pytorch-memonger is an open-source Python library designed for sublinear memory optimization in deep learning models built with PyTorch. It re-implements the technique described in "Training Deep Nets with Sublinear Memory Cost," significantly reducing the memory required for backward passes from O(N) to O(sqrt(N)). This optimization is particularly beneficial for training large sequential models or using larger batch sizes that would otherwise exceed available GPU memory. By replacing `nn.Sequential` with `memonger.SublinearSequential`, developers can easily integrate this memory-saving technique. The tool also addresses challenges with non-deterministic layers like BatchNorm and Dropout by re-scaling momentum and memorizing random number generator states, respectively, ensuring reliable performance.

sunnypilot is an open-source driver assistance system, forked from comma.ai's openpilot, designed to provide a unique driving experience across more than 350 supported car makes and models. It achieves this by modifying behaviors of driving assist engagements, all while adhering as closely as possible to comma.ai's safety policies. The system logs various data, including road-facing camera, CAN, GPS, IMU, and thermal sensors, with optional opt-in for driver-facing camera and microphone. Users can join the community forum for support and installation instructions, and documentation is available for features and FAQs. sunnypilot is released under the MIT License, with portions derived from openpilot.

UniFace is a lightweight, production-ready, and open-source face analysis library built on ONNX Runtime, offering a comprehensive suite of features for processing facial data. It includes advanced capabilities such as RetinaFace, SCRFD, YOLOv5-Face, and YOLOv8-Face for face detection with 5-point landmarks, and various models like AdaFace, ArcFace, and MobileFace for face recognition embeddings. The library also provides multi-object face tracking with BYTETracker, 106-point facial landmark localization, and semantic face parsing with BiSeNet. Additional features include trimap-free portrait matting, real-time gaze and head pose estimation, and attribute analysis for age, gender, race, and emotion. For privacy, it offers anti-spoofing and face anonymization with multiple blur methods. UniFace supports hardware acceleration on ARM64 (Apple Silicon), CUDA (NVIDIA), and CPU, making it versatile for different deployment environments.

U-2-Net is an open-source project that provides an implementation of the U^2-Net architecture, specifically designed for salient object detection. This deep learning model employs a nested U-structure, which allows for more profound analysis of image features and significantly enhances its ability to accurately identify and segment salient objects. The architecture is known for achieving improved performance in various computer vision tasks related to object detection and segmentation. It is a valuable resource for researchers and engineers working in computer vision, offering a robust framework for developing and experimenting with advanced image processing techniques.

torch-points3d is a comprehensive PyTorch framework designed for deep learning on point clouds, offering a robust environment for researchers and developers. It facilitates the implementation and experimentation of deep learning models for various 3D data analysis tasks. The framework is built upon PyTorch Geometric and Facebook Hydra, ensuring reproducibility and ease of model construction. It supports a high-level API to democratize deep learning on point clouds and includes implementations of state-of-the-art models like PointNet, PointNet++, RandLA-Net, and KPConv. Users can leverage it for tasks such as classification, segmentation, object detection, panoptic segmentation, and registration, with support for datasets like ScanNet, S3DIS, and ModelNet.

Vehicle-detection is an open-source project focused on implementing vehicle detection using machine learning and computer vision. It leverages several key techniques, including Linear Support Vector Machines (SVM) for classification, Histogram of Oriented Gradients (HOG) for feature extraction, color space conversion for image processing, and a sliding window approach to scan images for vehicles. This project was originally developed as part of Udacity's Self-Driving Car Engineer Nanodegree, indicating its practical application in autonomous vehicle technology research and development. It serves as a valuable resource for students and developers interested in computer vision and self-driving car applications.

VLFeat is an open-source library designed for computer vision algorithms, specializing in image understanding and local feature extraction and matching. Written in C for efficiency and compatibility, it offers MATLAB interfaces for ease of use and detailed documentation. The library supports popular algorithms including Fisher Vector, VLAD, SIFT, MSER, k-means, hierarchical k-means, agglomerative information bottleneck, SLIC superpixels, quick shift superpixels, and large-scale SVM training. It is compatible with Windows, Mac OS X, and Linux, making it a versatile tool for developers and researchers in the computer vision domain.

Monolith is an AI software platform designed for engineering product development, enabling engineers to build self-learning models that predict design performance. The platform helps reduce the need for extensive physical testing, accelerates learning from data, and ultimately improves product quality and time-to-market. It features an intuitive AI user interface with a notebook, built specifically for domain experts, and utilizes unique AI algorithms tailored for engineering applications. Monolith offers an enterprise SaaS cloud platform for large data and high-performance computing, alongside expert AI consulting to guide adoption and ensure success. Key modules include Test Data Validation, Test Plan Optimisation, and System Calibration, all aimed at streamlining engineering workflows and enhancing precision.

Mithrl is a scientific decision engine designed for enterprise R&D in the biomedical field. It significantly accelerates R&D cycles, reducing the time between experiments from weeks to an average of 32 times faster. The platform provides auditable, biologically grounded insights essential for high-stakes decisions. Mithrl helps researchers analyze omics data, contextualize findings with scientific rigor, and derive actionable insights. Key capabilities include identifying novel drug targets, discovering new biomarkers, interpreting results within a biological context, and decoding mechanisms of action, all by combining experimental data with Mithrl’s knowledge graphs.

AtlasNet is a sophisticated tool designed for learning 3D surface generation, leveraging a unique "papier-mâché" approach. This network is capable of synthesizing a complete mesh, including both point cloud data and connectivity, from various inputs such as low-resolution point clouds or even 2D images. It's an implementation of the research paper "AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation" presented at CVPR 2018. The repository provides source codes, installation instructions, and usage examples for both demo and training purposes. It supports Python 3.6, PyTorch, Pymesh, and Cuda 10.1, making it suitable for researchers and developers in the field of 3D deep learning and computer vision.

SaferPLACES is an AI-based digital twin solution designed for comprehensive flood risk intelligence. It integrates high-resolution geospatial, satellite, and climate data from sources like Google Earth Engine and Copernicus with advanced AI and physical-based models. This cloud-based platform provides real-time flood mapping and detailed flood hazard and damage assessments. SaferPLACES offers two cutting-edge solutions: a Cloud Web SaaS platform for effective climate adaptation planning and a tailored API for granular flood and climate risk data, ideal for parametric insurance and transparent risk disclosure. It empowers users to run computationally intensive simulations faster and more cost-effectively, supporting both private and public entities in identifying and mitigating current and future flood risks.

PianoRoll is a dedicated platform designed for piano enthusiasts to enhance their practice efforts. Users can practice, record their sessions or performances, and share them with a community of fellow piano lovers. The platform supports uploading previously recorded MIDI files and also offers a direct MIDI recorder. It emphasizes user ownership of content, stating that users retain full copyright of their work. Additionally, PianoRoll provides an option for users to donate their data to science for open-source datasets, contributing to research in music understanding and education, while strongly opposing unethical AI practices.