obss/sahi

Features

• Object Detection - Detects small objects in high-resolution images by dividing them into overlapping patches to avoid detail loss from downscaling.
• Image Tiling - Divides high-resolution images into smaller overlapping tiles to preserve object scale and avoid edge truncation.
• Image Slicing Pipelines - Implements a pipeline for dividing high-resolution images into smaller overlapping patches to improve small-object detection.
• Average Precision Calculators - Calculates class-wise average precision and average recall for standardized model benchmarking.
• Sliced - Splits large images into smaller overlapping patches to detect small objects that standard downscaling typically misses.
• Computer Vision Pipelines - Implements a standardized workflow for managing object detection and segmentation across various ML frameworks and backends.
• Machine Learning Model APIs - Provides a standardized programming interface to load and run predictions across different machine learning frameworks.
• Computer Vision Inference - Executes object detection and instance segmentation models on images and videos using a standardized processing pipeline.
• Model-Agnostic Inference Wrappers - Runs object detection and instance segmentation using various model frameworks via standard or sliced inference.
• Slicing Pipeline Integrations - Integrates diverse model backends into a unified pipeline to standardize high-resolution image processing.
• Model Performance Evaluators - Provides tools to quantify model accuracy and reliability by comparing predictions against ground truth labels.
• Tiled Inference Techniques - Splits large images and labels into overlapping tiles to maintain resolution during inference.
• Coordinate-Aware Slicing - Partitions high-resolution images and their annotations while maintaining accurate label mappings for each tile.
• Inference Batching - Groups multiple image slices into single processing units to maximize hardware throughput during inference.
• Unified Model Wrappers - Provides a unified programmatic interface that standardizes diverse machine learning framework APIs.
• Global Coordinate Remapping - Translates local slice coordinates back into the original image global coordinate system during result merging.
• Annotation Format Exporters - Provides tools to transform internal model predictions into standardized annotation formats for use with external analysis tools.
• Non-Maximum Suppression - Filters overlapping detections across slice boundaries using non-maximum suppression to eliminate redundant predictions.
• Frame Skipping Optimizations - Implements frame skipping optimizations to increase the processing speed of object detection across video streams.
• Detection Visualization - Converts sliced inference predictions into formats compatible with visual exploration and analysis tools.
• Batch Inference Pipelines - Implements a pipeline that groups multiple image slices into single model passes to maximize hardware throughput.
• Inference Execution - Executes a detection model on a full image without partitioning it into slices.
• Object Detection Evaluators - Calculates precision and recall metrics and generates error analysis plots by comparing predictions against ground truth.
• Prediction Visualization - Connects with visualization tools to inspect model predictions and ground truth data for large datasets.
• Classification Error Analysis - Generates performance breakdown plots by object size and error type to identify localization and classification failures.
• Intersection Over Union Analysis - Analyzes multiple detection result sets against ground-truth data using intersection-over-union thresholds.
• Video Object Segmentations - Processes video files by applying object detection and segmentation across frame sequences to track instances.
• Inference-Optimized Tuning - Optimizes object detection models specifically for sliced-inference pipelines to improve average precision.
• COCO Dataset Management - Provides a comprehensive toolkit for slicing, filtering, and converting COCO formatted annotations.
• COCO Dataset Processing - Processes COCO datasets through image slicing, annotation filtering, and format conversion to YOLO.
• COCO Dataset Toolkits - Offers a dedicated toolkit for slicing, filtering, and converting COCO formatted annotations.
• Annotation Format Translators - Transforms object detection annotations between standard schemas, specifically converting COCO format to YOLO.
• Image Processing Batchers - Runs inference across folders of images and automatically exports the resulting detections.
• Postprocessing Acceleration - Speeds up non-maximum suppression and merging by selecting high-performance backends to maximize CPU and GPU efficiency.
• Inference Workflow Automations - Enables running sliced inference on image directories and launching results in a visual app via a single command.
• Command Line Model Inferences - Provides a command-line interface for executing object detection predictions and dataset operations.
• Model Failure Analyzers - Calculates class-wise precision and recall to identify and analyze prediction failure modes.
• Computer Vision - Framework for sliced inference on large images.
• Deep Learning and Computer Vision - Library for sliced inference on large images.
• Image Processing and Manipulation - Performs sliced inference on large images for small object detection.

Star history