How mirapurkrabek/bboxmaskpose works
Project Type and Purpose
• This repository is an open-source research codebase, extending the MMPose 2.0 framework to implement BBox-Mask-Pose: a seamless loop of detection, segmentation, and pose estimation. It provides training and inference pipelines, model definitions, data loaders, and utilities to reproduce the MaskPose and RTMDet results from the paper.
File Structure
• Top-level package “mmpose”: core code implemented in subpackages (apis, codecs, configs, datasets, engine, evaluation, models, structures, utils, visualization).
• Configs: user-facing directory of model, dataset, and runtime configurations.
• Requirements & setup: external dependencies and installation helpers.
• Patterns: registry-driven builders, hook-based engine, modular data pipeline (datasets + transforms), plugin architecture for models/losses/heads.
README for Additional Insights
• Extensions over MMPose: multi-dataset ViTPose training, MaskBackground augmentation, integration of RTMDet from MMDetection.
• External dependencies: mmengine, mmcv, mmdet.
• Two primary workflows: training (config → engine → datasets/models → hooks/optimizers → evaluation) and inference (inferencer APIs → pre/post-processing → visualization).