thtrieu/darkflow
Darkflow
Darkflow est un framework de détection d'objets et un pipeline de vision par ordinateur qui fournit une interface programmatique pour effectuer de l'analyse d'image et de l'identification d'objets en temps réel. Il fonctionne comme un outil pour charger des poids, affiner des modèles et exécuter des inférences sur des images statiques et des flux vidéo.
Le projet sert de convertisseur qui traduit les configurations et les poids Darknet en graphes TensorFlow pour permettre le réentraînement et le déploiement. Il inclut un exportateur de modèles qui enregistre les graphes entraînés dans des fichiers protobuf portables pour une utilisation sur des appareils mobiles et natifs.
Le système couvre des capacités pour l'entraînement et l'ajustement de modèles de détection sur des jeux de données personnalisés, avec des points de contrôle de progression pour la reprise de l'entraînement. Il fournit également des outils pour la traduction de mappage de poids et le traitement de données d'image brutes via des opérations tensorielles pour produire des boîtes englobantes et des scores de confiance.
Features
- Object Detection - Identifies and locates objects within static images, returning bounding boxes, labels, and confidence scores.
- Computer Vision Pipelines - Provides an end-to-end computer vision pipeline for loading weights, fine-tuning models, and performing inference.
- Custom Model Training - Allows the import of existing weights to perform additional training on custom datasets for specific tasks.
- Detection Model Training - Trains and fine-tunes object detection models on custom datasets to classify specific objects.
- Model Inference APIs - Provides a Python-based interface to pass image arrays into detection models and receive structured results as dictionaries.
- Computer Vision Model Integration - Integrates pre-trained weights for object detection tasks using a programmatic Python interface.
- Object Detection Frameworks - Provides a comprehensive framework for real-time image analysis and object identification using deep learning.
- Cross-Framework Converters - Translates Darknet configurations and weights into TensorFlow graphs for retraining and deployment.
- Framework Converters - Translates Darknet models and weights into TensorFlow graphs to enable retraining and deployment.
- Real-Time Video Analysis - Analyzes live camera feeds or video files in real-time to identify objects and save processed output.
- Model Graph Conversion - Translates Darknet configuration files into TensorFlow computational graphs for execution and retraining.
- Portable Model Exporters - Converts trained graphs into portable protobuf files for mobile and native device deployment.
- Runtime Weight Loading - Implements runtime weight loading to align external binary weights with the target runtime's tensor dimensions.
- Training Checkpointers - Provides mechanisms for saving and resuming model training states to ensure session continuity.
- Weight Parameter Mapping - Provides utilities to map Darknet binary weight names and shapes to TensorFlow tensor variables.
- Model Export Formats - Saves trained graphs and weights into portable binary formats like Protobuf for deployment on mobile devices.
- Detection Model Exporters - Serializes trained detection models into portable Protobuf formats for deployment on mobile devices.
- Tensor Processing Pipelines - Implements a tensor processing pipeline that handles image batches as multi-dimensional arrays for GPU-accelerated inference.
- Mobile Model Deployment - Exports trained models into portable protobuf formats for execution on mobile devices.
- Computer Vision Libraries - TensorFlow port of the Darknet object detection framework.
- Computer Vision Models - Real-time object detection optimized for mobile device deployment.
- Object Detection - TensorFlow implementation of real-time object detection.
