repository·main·Indexed 12 days ago
https://github.com/facebookresearch/detectron2A modular PyTorch-based library for state-of-the-art detection and segmentation algorithms developed by Facebook AI Research. It provides tools for training and deploying models like Faster R-CNN, Mask R-CNN, and Keypoint R-CNN, with support for datasets including COCO, LVIS, Cityscapes, and Pascal VOC. The library includes a Model Zoo, lazy instantiation for configurations, and export capabilities for TorchScript and ONNX.
The tools/ directory contains several scripts for common tasks in the Detectron2 workflow:
train_net.py: An example training script for training built-in Detectron2 models. Refer to GETTING_STARTED.md for detailed usage.plain_train_net.py: A training script that implements a manual training loop instead of using the Trainer abstraction. This is intended for users who want more control ('hackers').benchmark.py: Benchmarks training, inference, or data loading speeds.analyze_model.py: Analyzes model FLOPs, parameters, and activations.visualize_json_results.py: Visualizes JSON results from COCOEvalutor or LVISEvaluator.visualize_data.py: Visualizes ground truth annotations or augmented training data.TridentNet (Scale-Aware Trident Networks) is an object detection architecture designed to generate scale-specific feature maps with uniform representational power. It uses a parallel multi-branch architecture where branches share transformation parameters but utilize different receptive fields.
This repository implements TridentNet-Fast, a fast approximation version that provides significant performance improvements without increasing parameters or computational cost.
detectron2.utils.visualizer module provides tools for visualizing model predictions (such as bounding boxes, segmentation masks, and keypoints) on images. It is the primary interface for creating visual representations of detection results.detectron2.utils.events module is used for managing and tracking events during training or inference, often used for logging metrics or triggering specific behaviors at certain intervals.detectron2.utils.analysis module contains tools for analyzing model performance, such as computing metrics or visualizing error patterns.The detectron2/export directory provides tools to prepare trained Detectron2 models for deployment in production environments. Currently, the export functionality supports the following formats:
For detailed usage instructions and tutorials, refer to the official deployment documentation.
detectron2.utils.colormap module provides utilities for generating and managing colormaps, typically used to assign distinct colors to different object classes in segmentation masks.detectron2.utils.memory module provides utilities to monitor or manage memory usage, which is critical when working with large models and high-resolution images on limited GPU memory.The DensePose Model Zoo provides various pre-trained models categorized by their architecture, training schedule, and feature sets. Users can choose models based on their specific needs for accuracy (AP), inference speed, or additional features like confidence estimation.
WC1 or WC2 types).WC1M or WC2M types).