open-mmlab/mmcv

Features

• Computer Vision Libraries - Acts as a foundation library for computer vision, providing high-performance operators, building blocks, and data pipelines.
• Custom Registrations - Allows registering user-defined modules and operators into the building system for configuration-based instantiation.
• Non-Maximum Suppression - Implements non-maximum suppression to filter overlapping bounding box predictions based on confidence scores.
• Computer Vision Preprocessing - Offers a comprehensive toolkit for preparing visual data via geometric and photometric transformations for ML tasks.
• Convolutional Architectures - Constructs convolution architectures for feature extraction using 2D, 3D, and adaptive weight standardization.
• Convolutional Neural Network Frameworks - Provides a specialized framework for creating and training convolutional neural networks using configurable layers.
• Spatial Image Padding - Provides utilities to pad image tensors for uniform dimensions during batch processing.
• Data Augmentation Pipelines - Organizes sequences of preprocessing and augmentation operations to prepare input data for machine learning.
• Deep Learning Architectures - Provides a configuration-driven framework for constructing multi-layered convolutional neural network architectures.
• Deep Learning Vision Operators - Ships a collection of optimized kernels for deformable convolutions, region pooling, and non-maximum suppression.
• GPU Accelerated Computer Vision - Leverages graphics hardware to provide high-performance acceleration for real-time image analysis and manipulation.
• Image Data Preprocessing - Provides a toolkit for preparing raw image and video data for deep learning model consumption.
• Neural Network Layers - Instantiates activation, normalization, padding, and upsampling layers via a configuration-driven interface.
• Bundled Convolutional Layers - Combines convolution, normalization, and activation layers into a single module using configurable parameters.
• Deformable Convolutions - Provides high-performance deformable convolution layers that adapt their receptive field using learnable offsets.
• Rectangular Region Extraction - Extracts rectangular regions from images based on specified coordinate bounds.
• Image Flipping - Implements horizontal and vertical mirroring of images for data augmentation.
• Image Rotation - Allows rotating images by specified angles with options to crop or expand the canvas.
• Dimension Resizing - Adjusts pixel dimensions of images to fit layout constraints and optimize memory usage.
• Image File IO - Loads images from files or bytes and saves processed data to the filesystem.
• Image-to-Tensor Conversions - Converts processed raster image pixel data into numerical tensors suitable for deep learning models.
• Computer Vision Operator Acceleration - Provides hardware-accelerated kernels for complex vision operations like deformable convolutions and region pooling.
• Spatial-Level Transformations - Provides comprehensive geometric operations such as rotation, cropping, and warping for image data.
• Configuration-Driven Instantiations - Constructs network layers using dictionary specifications to enable dynamic swapping of layer types and parameters.
• Image Property Adjusters - Alters image brightness, contrast, hue, sharpness, and histogram distribution.
• Criss-Cross Attention - Implements a cross-shaped attention mechanism to capture long-range dependencies for semantic segmentation tasks.
• Border Feature Extraction - Samples and pools features along bounding box boundaries using bilinear interpolation to refine object detection.
• Region Alignment - Ensures spatial alignment during feature pooling for specific bounding boxes using interpolation.
• Region Pooling - Extracts fixed-size feature maps from variable-sized regions of interest using pooling operations.
• Corner Pooling - Extracts features from the corners of a region to improve object keypoint detection.
• Custom Data Transform Extensions - Provides a mechanism for defining and registering custom data transformation classes via a base class.
• Transform Sequence Assemblers - Assembles sequences of transforms for loading, augmenting, and packing image data into training pipelines.
• Intersection Over Union Calculators - Calculates the intersection over union (IoU) for bounding boxes in 2D, 3D, and bird's eye view.
• Model Complexity Analysis - Calculates the computational complexity and resource requirements of neural network architectures to optimize inference.
• Stochastic Augmentation Pipelines - Executes image transformations or pipeline choices based on defined probabilities for dataset augmentation.
• Point-to-Voxel Conversion - Transforms raw 3D point coordinates into a voxelized grid representation for object detection encoders.
• Point Cloud Processing - Provides libraries for manipulating, filtering, and voxelizing 3D point cloud data for object detection.
• Sampling - Selects representative points from 3D point clouds using furthest point sampling and ball queries.
• Custom Dataset Loading - Implements logic for loading raw image files and corresponding annotations from disk into tensors for training.
• Annotation Rendering - Renders images with corresponding labels or bounding boxes to verify data quality and model predictions.
• Localization Data Loading - Provides utilities to load images paired with spatial annotations like bounding boxes and masks for model training.
• Color Space Conversions - Transforms images between diverse color models including RGB, BGR, HSV, HLS, YCbCr, and grayscale.
• Tensor-to-Image Reconstructions - Converts numerical tensors back into grayscale or color image files for visualization and storage.
• Kernel Fusion Operations - Fuses convolution and batch normalization layers into single operations to reduce inference overhead.
• Sparse Convolution Operators - Implements high-performance sparse convolutions and pooling operators for high-dimensional spatial data.
• Transformation Broadcasting - Applies a single transformation to multiple data fields or targets, such as images and bounding boxes, simultaneously.
• Computer Vision Frameworks - Foundational library for image and video processing.

Star history