Projects

### Overview: This comprehensive course delves into the intricacies of Convolutional Neural Networks (CNNs) and their applications in visual recognition tasks. Focusing on image classification, this course provides a deep dive into the latest research in computer vision, emphasizing hands-on learning and practical engineering techniques. ### Key Features: - Implementing, training, and debugging neural networks. - Gaining a detailed understanding of cutting-edge research in computer vision. - Hands-on assignments and a final course project involving a multi-million parameter convolutional neural network. - Training on the largest image classification dataset (ImageNet). ### Tech Stack: - Jupyter Notebook ### Why It Matters: This course is essential for developers and researchers interested in advancing their skills in computer vision, particularly in the areas of image classification and recognition. By mastering the techniques taught in this course, participants can contribute to the development of state-of-the-art visual recognition systems and applications in fields such as medicine, drones, and self-driving cars. ### Repository: khanhnamle1994/computer-vision ### Primary Language: Jupyter Notebook ### GitHub Stars: 510

### Overview: GFNet is a transformer-style architecture designed for image classification tasks, achieving state-of-the-art performance on ImageNet. It utilizes a 2D discrete Fourier transform, element-wise multiplications, and inverse Fourier transform to replace self-attention and spatial MLP, making it computationally efficient and highly effective for long-term spatial dependencies. ### Key Features: - **Transformer Architecture**: GFNet is a transformer-based model that processes images using a combination of 2D discrete Fourier transform, element-wise multiplications, and inverse Fourier transform. - **Efficiency**: It significantly reduces computational complexity compared to self-attention and spatial MLP, making it suitable for high-resolution images. - **Pretrained Models**: The repository includes pretrained GFNet models for ImageNet, available on various cloud storage platforms. ### Tech Stack: - **Language**: Jupyter Notebook - **GitHub Stars**: 510 ### Why It Matters: GFNet offers a computationally efficient alternative to traditional image classification methods, making it particularly useful for applications requiring high accuracy and low computational resources. Its ability to process high-resolution images efficiently makes it a valuable tool in various domains, including computer vision and image processing.

Overview: This project focuses on predicting dense depth maps from sparse and noisy LiDAR frames using RGB guidance, aiming to improve depth completion tasks. The work was presented at the 2019 Machine Vision Applications (MVA) conference and achieved the 1st place on the KITTI depth completion benchmark. Key features: - Predicting dense depth maps from sparse and noisy LiDAR frames - Guided by RGB images for accurate depth completion - Uses uncertainty to combine multiple sensor data - Ranks first on the KITTI depth completion benchmark without additional data or postprocessing Tech stack: Python Why it matters: This project addresses a significant challenge in depth completion tasks, providing a robust method for generating accurate depth maps from sparse LiDAR data, which is crucial for applications requiring high precision in depth perception. Repository: wvangansbeke/Sparse-Depth-Completion Primary language: Python GitHub stars: 510 GitHub description: Predict dense depth maps from sparse and noisy LiDAR frames guided by RGB images. (Ranked 1st place on KITTI) [MVA 2019] License: Creative Commons Attribution-NonCommercial license Introduction: This project combines global and local information from LiDAR and RGB images to complete sparse and noisy LiDAR frames, achieving the 1st place on the KITTI depth completion benchmark. The method uses confidence maps for late fusion, making it effective without additional data or postprocessing. Requirements: Python 3.7 Dataset: The project uses the

### Overview: BertSimilarity 是一个基于 Google BERT 模型的句子相似度计算工具，用于文本和语义相似度的计算。通过将句子拼接并使用 BERT 进行编码，该工具能够快速准确地计算句子之间的相似度。 ### Key features: - **Sentence Similarity Calculation**: 计算两个句子的相似度。 - **Google BERT Model**: 使用 Google 的 BERT 模型进行句子编码。 - **Python Implementation**: 代码基于 TensorFlow 1.x 实现。 ### Tech stack: Python ### Why it matters: BertSimilarity 提供了一种高效且准确的方法来计算句子的相似度，适用于自然语言处理和文本分析任务。

### Overview: PersFormer is an end-to-end monocular 3D lane detector that leverages a novel Transformer-based spatial feature transformation module. This model generates BEV features by attending to related front-view local regions with camera parameters as a reference, enhancing feature consistency and benefiting from multi-task learning. ### Key Features: - Perspective Transformer Module - Unified 2D/3D Anchor Design - Auxiliary Task for Simultaneous 2D/3D Lane Detection ### Tech Stack: Python ### Why It Matters: PersFormer stands out in 3D lane detection by providing a unified approach that enhances feature consistency and leverages multi-task learning, leading to superior performance on the OpenLane benchmark. This model has been recognized with an ECCV 2022 Oral presentation and is recommended for detailed analysis and further research. ### Repository: Repository: [OpenDriveLab/PersFormer_3DLane](https://github.com/OpenDriveLab/PersFormer_3DLane) Stars: 509 Description: [ECCV 2022 Oral] Perspective Transformer on 3D Lane Detection ### Changelog: - 2022-11-3: Fixed evaluation pipeline bug and uploaded PersFormer model on openlaneV1.1 for reproducibility. - 2022-9-27: Updated evaluation metrics, pruned gt points by visibility, and supported Gen-LaneNet on OpenLane. - 2022-5-9: Compared PersFormer on [ONCES_3

Overview: Population Based Augmentation (PBA) is an algorithm designed to quickly and efficiently learn data augmentation functions for neural network training. It achieves state-of-the-art results on CIFAR with one thousand times less compute, making it suitable for researchers and practitioners to learn new augmentation policies using a single workstation GPU. Key features: - Efficient learning of augmentation schedules - Reduced compute time compared to state-of-the-art methods - Visualizable augmentation strategy Tech stack: TensorFlow and Python Why it matters: PBA enables researchers and practitioners to effectively learn new augmentation policies using a single workstation GPU, significantly reducing the computational requirements for training neural networks.

Overview: Vigil is a Python library and REST API designed to assess Large Language Model (LLM) prompts and responses for potential security threats, such as prompt injections and jailbreaks. This repository includes detection signatures and datasets for self-hosting. Key features: - Analyze LLM prompts for common injections and risky inputs - Use Vigil as a Python library or REST API - Scanners are modular and easily extensible - Evaluate detections and pipelines with Vigil-Eval (coming soon) - Available scan modules: Vector database / text similarity, YARA heuristics, Transformer model, Prompt-response similarity, Canary Tokens, Sentiment analysis, Paraphrasing - Supports local embeddings and OpenAI - Signatures and embeddings for common attacks - Custom detections via YARA signatures Tech stack: Python Why it matters: Vigil is a critical tool for developers and security professionals to identify and mitigate potential security threats in LLM prompts, ensuring the integrity and safety of their applications. Repository: deadbits/vigil-llm GitHub stars: 509

Overview: This repository contains translated and compiled code from the book "Programming Computer Vision with Python," providing practical examples for developers. Key features: - Translated Python code from the book - Compilation of examples from the book - Code still under development Tech stack: Python Why it matters: This repository serves as a practical resource for developers looking to understand and implement computer vision concepts using Python, offering insights and code snippets from a well-regarded book. Repository: willard-yuan/pcv-book-code Primary language: Python GitHub stars: 509 GitHub description: :book:Python计算机视觉中译本实例代码 README excerpt:

### Overview: `bpycv` is a collection of computer vision and deep learning utilities designed specifically for Blender. It simplifies the process of generating instance annotations, depth maps, and 6D pose ground truth using a single line of code. This tool is particularly useful for researchers and developers looking to create synthetic datasets for training and testing purposes. ### Key Features: - **Render Instance Annotations**: Generate semantic, instance segmentation, and panoptic segmentation annotations for images. - **6DoF Pose Ground Truth**: Provide accurate 6D pose ground truth for 3D reconstruction. - **Pre-defined Domain Randomization**: Offers options for randomizing lighting and backgrounds, as well as distractors from ShapeNet. - **Easy Installation and Demo Running**: Supports installation via Docker, making it straightforward to set up and use. - **Python Codebase**: Provides a Python-based codebase for building synthetic datasets, enhancing flexibility and ease of development. - **Conversion to Cityscapes Format**: Converts the generated data into the Cityscapes annotation format for compatibility. - **Docker Support**: Available with a Dockerfile for easy deployment. - **No Complicated Packaging**: Simplifies the process of packaging and debugging, making it ideal for both research and development. ### Tech Stack: - **Python**: The primary programming language used for developing and running the utility. ### Why It Matters: `bpycv` streamlines the process of creating synthetic datasets, making it easier to train models on large datasets. This utility is particularly valuable for researchers and developers looking to improve the accuracy and diversity of their datasets

### Overview: This repository contains the code for the third edition of "Transformers for Natural Language Processing and Computer Vision," authored by Denis Rothman. The book explores the advancements in Large Language Models (LLMs), including Hugging Face, OpenAI, and Google Vertex AI, and their applications in Natural Language Processing (NLP) and Computer Vision (CV). ### Key Features: - Implementing Generative AI and Large Language Models (LLMs) - Pretraining and fine-tuning LLMs - Working with multiple platforms, such as Hugging Face, OpenAI, and Google Vertex AI - Preprocessing language data using tokenizers - Visualizing transformer model activity with BertViz, LIME, and SHAP - Chaining models for better results - Fine-tuning OpenAI GPT models - Investigating the role of tokenizers in shaping transformer models - Leveraging LLM embeddings for alternative fine-tuning - Exploring vision transformers with CLIP, DALL-E 2, DALL-E 3, and GPT-4V - Creating and implementing cross-platform chained models - Investigating the risks in large language models - Beyond text: exploring vision transformers in the dawn of revolutionary AI - Transcending the image-text boundary with Stable Diffusion - Training vision models without coding with Hugging Face AutoTrain - On the road to functional AGI with HuggingGPT and its peers - Beyond human-designed prompts with generative ideation ### Tech Stack: Jupyter Notebook ### Why it

### Overview: TP-GAN is an official implementation of the Tensorflow framework for the ICCV17 paper "Beyond Face Rotation: Global and Local Perception GAN for Photorealistic and Identity Preserving Frontal View Synthesis". This project aims to recover a frontal face image of the same person from a single face image under any poses. ### Key features: - **Frontal Image Synthesis**: TP-GAN synthesizes frontal face images from single face images under any poses. - **Global and Local Perception**: The model uses a combination of global and local perception to achieve photorealistic and identity-preserving synthesis. - **MultiPIE Dataset**: The project includes synthesized testing images from the MultiPIE dataset, which can be accessed for further analysis. ### Tech stack: Python ### Why it matters: TP-GAN is a significant contribution to the field of photorealistic face synthesis, offering a practical solution for generating frontal views from existing face images. Its ability to handle various poses and maintain identity is particularly noteworthy, making it valuable for applications in augmented reality, virtual reality, and facial recognition systems. ### Repository: HRLTY/TP-GAN - **Primary language**: Python - **GitHub stars**: 508 - **GitHub description**: Official TP-GAN Tensorflow implementation for the ICCV17 paper "[Beyond Face Rotation: Global and Local Perception GAN for Photorealistic and Identity Preserving Frontal View Synthesis](http://openaccess.thecvf.com/content_ICCV_2017/papers/Huang_Beyond_Face

### Overview: This repository is a curated collection of Open Information Extraction (OIE) resources, including papers, code, data, and applications. It serves as a comprehensive resource for researchers and developers interested in the field of OIE. ### Key features: - **Papers Sorted Chronologically**: Organized by year, providing a chronological view of the evolution of OIE research. - **Papers Grouped by Category**: - **Surveys**: Comprehensive reviews and analyses of OIE systems. - **Evaluation**: Metrics and benchmarks for evaluating OIE systems. - **OIE for Downstream Applications**: Applications of OIE in specific domains like question answering, slot filling, event extraction, etc. - **OIE in Different Languages**: Resources for OIE systems in various languages such as German, Portuguese, Spanish, Chinese, Persian, Italian, and Indonesian. - **Supervised OIE**: Resources for supervised OIE systems. - **Canonicalization of OIE**: Techniques for canonicalizing OIE data. - **Slides**: Presentations and talks related to OIE. - **Code**: Implementations and code samples for OIE systems. - **Data**: OIE corpora and resources derived from OIE output. - **PhD Theses**: Research papers from PhD theses in the field of OIE. - **Demos**: Demonstrations and examples of OIE systems in action. ### Tech Stack: The repository is built using Markdown, GitHub Pages, and various tools for data organization and documentation. ### Why it Matters: This curated collection

### Overview: Ultralytics YOLO Flutter is an official plugin for running YOLO models in Flutter apps on iOS and Android. It supports detection, instance segmentation, semantic segmentation, depth estimation, classification, pose, and OBB on both platforms, with two simple entry points: `YOLO` for single-image inference and `YOLOView` for real-time camera inference. ### Key Features: - Official Ultralytics plugin for Flutter integration. - One Dart API for Android and iOS. - Metadata-first model loading with official model download and caching. - Real-time camera inference and single-image inference. - Controls for thresholds, accelerator selection, and result streaming. - YOLO26 and YOLO11 model families supported. ### Tech Stack: Dart ### Why It Matters: Ultralytics YOLO Flutter simplifies the integration of YOLO models in Flutter apps, making it easy to deploy real-time inference tasks on Android and iOS. The plugin supports a wide range of vision tasks and provides a straightforward way to use official models or your own exported models.

Overview: *wego* is an implementation of word embeddings models in Go, supporting Word2Vec, GloVe, and LexVec models, and providing nearest neighbor search tools for word vectors. Key features: - Word2Vec: Distributed Representations of Words and Phrases - GloVe: Global Vectors for Word Representation - LexVec: Matrix Factorization using Window Sampling and Negative Sampling Tech stack: Go Why it matters: *wego* provides a practical and efficient way to generate word embeddings, making it useful for various applications in natural language processing, such as text classification, recommendation systems, and more. Repository: ynqa/wego GitHub stars: 508 GitHub description: Word Embeddings in Go!

Overview: BETO is a Spanish version of the BERT model, trained on a large Spanish corpus. It is similar in size to a BERT-Base and uses Whole Word Masking. The repository includes checkpoints for both cased and uncased versions, as well as results for Spanish benchmarks. Key features: - Spanish BERT model - Cased and uncased versions - Trained on a big Spanish corpus - Size similar to BERT-Base - Uses Whole Word Masking technique - Available on HuggingFace Model Repository Tech stack: TensorFlow and PyTorch checkpoints Why it matters: BETO is useful for Spanish language tasks such as part-of-speech tagging, named entity recognition, and machine learning documentation, providing performance comparable to Multilingual BERT and other non-BERT methods.

Overview: 这是一个基于BERT和CRF的韩语命名实体识别模型的PyTorch实现。 Key features: - 使用BERT和CRF进行韩语命名实体识别 - 支持多种实体类型，如人名、地名和组织名 - 支持多GPU和TPU训练 Tech stack: PyTorch v1.2 / Python 3.x Why it matters: 该模型为韩语命名实体识别提供了高效、准确的解决方案，适用于各种文本处理任务。

Overview: ModelScope Cookbook is a comprehensive guide for developers, providing a structured learning path from model selection to practical application development. It covers model inference, data preparation, fine-tuning, evaluation, and application creation using tools like EvalScope, ms-swift, DiffSynth, and Ollama. Key features: - Model selection - Inference - Data preparation - Fine-tuning - Evaluation - Application development Tech stack: HTML Why it matters: ModelScope Cookbook helps developers bridge the gap between open-source models and practical AI applications, making it easier to build and deploy models in real-world scenarios.

### Overview: This repository collects and surveys papers that integrate Knowledge Graphs (KGs) and Multi-Modal Learning, focusing on research in two principal aspects: **KG-driven Multi-Modal (KG4MM) learning**, where KGs support multi-modal tasks, and **Multi-Modal Knowledge Graph (MM4KG)**, which extends KG studies into the MMKG realm. ### Key Features: - **KG4MM Learning**: Papers that leverage KGs to enhance multi-modal learning tasks. - **MM4KG**: Papers that explore the integration of KGs into the MMKG domain, expanding the scope of knowledge graph studies. ### Tech Stack: The repository uses Markdown, GitHub, and various LaTeX packages for documentation and publication. ### Why It Matters: This comprehensive survey provides a valuable resource for researchers and practitioners interested in the intersection of KGs and multi-modal learning, offering insights into the latest advancements and future directions in the field.

### Overview: Discover the cutting-edge techniques and papers in the field of Radiance Field-based 3D Editing. This curated list is your gateway to understanding and mastering the latest advancements in 3D computer vision and graphics. ### Key Features: - **NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis** - ECCV 2020 Oral - **3D Gaussian Splatting for Real-Time Radiance Field Rendering** - SIGGRAPH 2023 Best Paper - **Neural Fields in Visual Computing and Beyond** - Eurographics 2021 - **A Revisit of Shape Editing Techniques: from the Geometric to the Neural Viewpoint** - JCST 2021 - **Advances in Neural Rendering** - Eurographics 2022 - **Multimodal Image Synthesis and Editing: The Generative AI Era** - TPAMI 2023 - **Recent Advances in 3D Gaussian Splatting** - CVM 2024 - **3D Gaussian as a New Vision Era: A Survey** - TVCG 2025 - **3D Gaussian Splatting: Survey, Technologies, Challenges, and Opportunities** - IEEE TCSVT 2025 - **Editing Implicit and Explicit Representations of Radiance Fields: A Survey** - MVA 2025 - **A Survey on 3D Gaussian Splatting Applications: Segmentation, Editing, and Generation** - TPAMI

Overview: Caura, formerly MemClaw, is an open-source shared memory layer designed for multi-tenant, multi-agent AI fleets. It ensures governed, searchable, and self-improving memory, enabling agents to learn from each other rather than repeating mistakes. Key features: - Multi-tenant, multi-agent AI fleets - Governed, searchable, self-improving memory - One loop, three pillars: write, recall, compound - Built for fleets, with headroom for scalability Tech stack: Python Why it matters: Caura provides a shared, governed memory layer for AI agents, ensuring that agents learn from each other rather than repeating mistakes. This makes it suitable for multi-tenant, multi-agent environments, where agents can collaborate and improve together. Repository: caura-ai/caura GitHub stars: 500 GitHub description: Apache 2.0 licensed, open-source memory for multi-tenant, multi-agent AI fleets.