ultralytics/yolov3
Yolov3
This is a real-time object detection framework built on the YOLOv3 architecture, implemented in PyTorch. It provides a complete pipeline for identifying and localizing objects in images and video using a single neural network pass, combining a Darknet-53 backbone with multi-scale feature pyramids and anchor-based bounding box prediction.
The framework extends beyond basic detection to include instance segmentation, human pose estimation, and multi-object tracking across video frames. It offers a model export toolkit that converts trained models through ONNX to CoreML, TensorFlow Lite, and TensorRT formats for cross-platform deployment. A PyTorch training pipeline supports custom model training on user-provided datasets, with cloud-based experiment tracking through Weights & Biases and Roboflow for logging metrics and managing runs.
Training capabilities include architecture selection among model variants, hyperparameter tuning for epochs, resolution, and batch size, and validation of precision, recall, and mean average precision. Data preparation tools support custom dataset labeling and assembly of annotated images for training.
Features
- YOLOv3 Frameworks - Implements the YOLOv3 architecture for real-time object detection in images and video.
- YOLOv3 Implementations - Provides a complete real-time object detection framework built on the YOLOv3 architecture.
- Anchor Box Systems - Predicts object locations relative to predefined anchor box shapes at each grid cell.
- Non-Maximum Suppression - Removes duplicate detections by keeping only the highest-confidence bounding box among overlapping predictions.
- Object Detection - Runs a pretrained neural network on an image to locate and classify objects within it.
- Real-Time Object Detection - Provides real-time object detection on images and video streams using a single neural network pass.
- Detection Model Training - Fine-tunes a pretrained detection model on a user-provided labeled dataset to recognize new object categories.
- Multi-Scale Feature Pyramids - Extracts object features at three spatial scales to detect both small and large objects.
- Darknet-53 Backbones - Provides the Darknet-53 backbone for hierarchical feature extraction in real-time detection.
- Object Detection - Runs inference on images and videos to locate and classify objects using trained detection models.
- PyTorch Training Frameworks - Provides a PyTorch training pipeline for fine-tuning YOLOv3 models on custom datasets.
- Performance Validators - Provides built-in validation of precision, recall, and mean average precision for trained detection models.
- Single Shot Detectors - Runs a single neural network pass per image to simultaneously predict bounding boxes and class probabilities.
- Human - Ships a human pose estimation model that detects body keypoints and reconstructs poses in images and video.
- Instance Segmentation Engines - Assigns a pixel-level mask to each detected object, separating it from the background and other objects.
- Video Stream Detections - Processes video frames sequentially to identify and track objects across a video stream.
- Detection Model Validation - Evaluates precision, recall, and mean average precision against a validation dataset.
- Experiment Tracking Integrations - Integrates with Weights & Biases and Roboflow for automated experiment tracking and dataset management.
- Model Architecture Selection - Selects between pretrained and scratch model variants to balance accuracy, speed, and memory.
- Hyperparameter Tuning - Adjusts epochs, resolution, batch size, and augmentation to improve model convergence and accuracy.
- Weights and Biases Training Logging - Streams training metrics to Weights & Biases for real-time visualization and experiment comparison.
- Object Tracking - Assigns a persistent ID to each detected object and follows its movement through a video sequence.
- Pose Estimation - Detects key body joints and limbs for each person in an image or video to reconstruct their pose.
- PyTorch-to-ONNX Pipelines - Converts trained models through ONNX to CoreML, TensorFlow Lite, and TensorRT formats.
- Image Augmentations - Applies random transformations like scaling, cropping, and color jittering to training images.
- Multi-Object Trackers - Assigns persistent IDs to detected objects and tracks their movement across video frames.
- Video Object Tracking - Assigns persistent IDs to detected objects and follows their movement across video frames.
- Model Export Formats - Exports trained models to ONNX, CoreML, TFLite, and TensorRT for production deployment.
- Multi-Format Exporters - Converts trained models to ONNX, CoreML, TFLite, and TensorRT for deployment across platforms.
- Computer Vision - Classic real-time object detection model.
- Computer Vision Libraries - YOLOv3 implementation for PyTorch and mobile export.
- Object Detection - Listed in the “Object Detection” section of the The Incredible Pytorch awesome list.
