repository·main·Indexed 27 days ago
https://github.com/open-mmlab/mmcvA foundational library for computer vision research within the OpenMMLab ecosystem. MMCV provides essential tools for image and video processing, visualization, CNN architectures, and high-performance CPU/CUDA operations. It includes modules for image IO and geometric transformations (mmcv.image), neural network layer construction and specialized convolutional modules (mmcv.cnn), and high-performance computer vision operations (mmcv.ops). The library is available in a full version with CUDA ops and a lightweight version called mmcv-lite.
MMCV is a fundamental library for computer vision that provides essential utilities including:
MMCV supports Linux, Windows, and macOS. It requires Python 3.7 or higher.
MMCV is a fundamental library for computer vision that provides core functionalities including:
MMCV is a foundational library for computer vision research that provides core functionalities including:
The mmcv.transforms module provides a suite of tools for data augmentation, loading, and processing in computer vision pipelines. It is organized into three main functional categories:
LoadImageFromFile and LoadAnnotations.Resize, RandomResize), cropping (CenterCrop), flipping (RandomFlip), normalization (Normalize), and tensor conversion (ToTensor, ImageToTensor).Compose for chaining operations, RandomApply or RandomChoice for stochastic augmentations, and KeyMapper for managing data keys.The mmcv.video module provides high-level functions for video manipulation:
concat_video: Concatenate multiple video files.convert_video: Convert video formats.cut_video: Cut/trim segments from a video.resize_video: Resize video dimensions.In MMCV, data preparation is decoupled from dataset construction. Data transforms are callable classes that accept a configuration during instantiation and process a data dictionary.
Key Conventions:
dict as input and returns a dict as output.Resize reads img) and may add or update other fields.Resize, Pad), the image size order is (width, height).img_shape, ori_shape, pad_shape), the order is (height, width).>>> import numpy as np
>>> from mmcv.transforms import Resize
>>>
>>> transform = Resize(scale=(224, 224))
>>> data_dict = {'img': np.random.rand(256, 256, 3)}
>>> data_dict = transform(data_dict)
>>> print(data_dict['img'].shape)
(224, 224, 3)MMCV is part of the OpenMMLab ecosystem. You can use other specialized toolboxes for various computer vision and deep learning tasks:
mmcv.ops module provides a wide range of high-performance computer vision operations, including specialized convolutions, pooling layers, attention mechanisms, and geometric utilities. These operations are optimized for efficiency and are commonly used in object detection, segmentation, and 3D perception tasks.The mmcv.video module includes several functions for handling optical flow data:
dequantize_flow: Dequantize optical flow.flow_from_bytes: Convert optical flow from bytes.flow_warp: Warp an image using optical flow.flowread: Read optical flow files.flowwrite: Write optical flow files.quantize_flow: Quantize optical flow.sparse_flow_from_bytes: Convert sparse optical flow from bytes.The mmcv.visualization module provides utilities for handling colors, visualizing images (including bounding boxes), and processing optical flow data.
Color: A class for color representation and manipulation.color_val: Utility for color values.imshow: Display an image.imshow_bboxes: Display an image with bounding boxes.imshow_det_bboxes: Display an image with detection bounding boxes.flow2rgb: Convert optical flow to RGB.flowshow: Display optical flow.make_color_wheel: Generate a color wheel.mmcv.ops module provides a wide range of high-performance computer vision operations, including specialized convolutional layers, pooling operations, attention mechanisms, and geometric utilities. These operations are often implemented with CUDA acceleration for efficiency in deep learning tasks like object detection, segmentation, and 3D perception.The MMCV custom ops C++ source code is organized into several functional directories:
common: Shared tools and kernels. Includes cuda (shared CUDA/HIP kernels), mps (experimental Metal Performance Shaders support), mlu (Cambricon device support), and utils (e.g., spconv kernels).parrots: Custom ops specifically for the Parrots deep learning framework.pytorch: PyTorch-specific implementations using pybind11. Subdirectories include:cuda: CUDA kernel launchers and C++ interfaces.cpu: CPU implementations.mlu: MLU kernel launchers.mps: MPS implementations and launchers.