MMHuman3D Release v0.11.0

Highlights

• Add ExPose inference
• Add PyMAF-X inference
• Support CLIFF

New Features

• Add ExPose inference
• Add PyMAF-X inference
• Support CLIFF
• Add reordered MANO convention
• Add mediapipe convention

Documentation

• Update SMC Documents

Bug Fixes

• Fix ExPose config error for inference
• Change flake8 url from gitlab to github
• Fix slice bugs for HumanData
• Fix bugs in ExPose inference
• Fix bugs in HumanImageDataset
• Fix img_res in configs

Many thanks to this release's contributors!
@oneScotch @yl-1993 @jiaqiAA @KHao123 @Zessay @caizhongang @ttxskk

MMHuman3D Release v0.10.0

Highlights

• Add webcam demo and real-time renderer
• Update dataloader to speed up training
• Add balanced MSE loss for imbalanced HMR training

New Algorithms

• Add balanced MSE loss for imbalanced HMR training (added by paper authors)

New Features

• Add persistent_workers in dataloader to reduce data time in training
• Add webcam demo and real-time renderer
• Support mmhuman3d installation without torch

Documentation

• Add smc file description
• Update install.md to address dependency conflicts
• Update GTA-Human, main readme, and install guide
• Update doc to avoid mmcls and mmtrack installation issues

Bug Fixes

• Fix error raise messages in estimate_smpl.py
• Fix ImportError caused by missing __init__.py
• Fix config not found when installed by 'pip install git+...'
• Fix GTA-Human config path error
• Fix bugs in guas1d config
• Fix bug for the real-time renderer

Many thanks to this release's contributors!
@ykk648 @haofanwang @woo1 @mingyuan-zhang @ttxskk @kimren227 @caizhongang @LazyBusyYang @yl-1993

MMHuman3D Release v0.9.0

Highlights

• Support SMPL-X estimation with ExPose for simultaneous recovery of face, hands and body
• Support new body model STAR
• Release of GTA-Human dataset with SPIN-FT (51.98 mm) and PARE-FT (46.84 mm) baselines! (Official)
• Refactor registration and improve performance of SPIN to 57.54 mm

* Values are PA-MPJPE on 3DPW test

New Algorithm

• Support training and evaluation of SMPL-X estimation method ExPose
• Add a general architecture for expressive human modelling
• Add data preparation and preprocessing tools
• Add evaluation metrics

New Body Models

• Support new body model STAR by adapting the official PyTorch implementation into MMHuman3D
• Support FLAME and MANO needed by ExPose

New Dataset

• Official release of our large-scale synthetic dataset GTA-Human that contains 1.4M SMPL annotations
• Add strong baselines SPIN-FT and PARE-FT baselines, which achieve 51.98 mm and 46.84 mm PA-MPJPE on 3DPW test respectively

Refactors

• Refactor registration
  • Allow more flexible registration configurations
  • Improve performance of SPIN to 57.54 mm PA-MPJPE on 3DPW test

Documentation

• Add documation for installing MMHuman3D on Windows

Bug Fixes

• Fix missing arguments in SMPLify pipeline
• Adjust import structure in core, data and utils, removing unnecessary dependency on PyTorch3D, and achieving up to 2.6x speed-up during initialization

Many thanks to this release's contributors!

@kristijanbartol @mingyuan-zhang @Coach257 @caizhongang @LazyBusyYang @ttxskk

MMHuman3D Release v0.8.0

Highlights

• Support SmoothNet (added by paper authors)
• Fix circular import and up to 2.5x speed up in module initialization
• Add documentations in Chinese

New Algorithms

• Support SmoothNet as a smoothing method, improving inference precision for image- and video-based models

Refactors

• Refactor estimate_smpl.py with compressed pose and shape to avoid possible shape errors and waste of space for multi-person visualization

Documentations

• Add Chinese version for 12 documentations

Bug Fixes

• Fix circular import in models, conventions, and cameras, and achieve up to 2.5x speed-up in module initialization by importing only the necessary builders
• Fix bugs in PARE configs
• Fix wrong type and value in convert_kps

Many thanks to this release's contributors!

@Naive-Bayes @ykk648 @juxuan27 @LazyBusyYang

MMHuman3D Release v0.7.0

Highlights

• Support PARE (better than the official implementation)
• Support DeciWatch (added by paper authors)
• Add GTA-Human HMR baseline (official release)
• Support saving inference results

New Algorithms

• Support PARE with 49.35 mm PA-MPJPE on 3DPW, better than official implementation (50.9 mm)
• Support DeciWatch for both smoothing and up to 10x inference speed-up, added by authors of the paper
• Official release of GTA-Human HMR baseline with 11.82 mm PA-MPJPE improvement on 3DPW than the original HMR

New Features

• Support saving inference results in HumanData format

Refactors

• Merge HMR's and SPIN's H36M converter so the same preprocessed file is used now

Bug Fixes

• Fix bugs in H36M and MPI-INF-3DHP data converters
• Fix typo in smpl visualization tool
• Fix in_ndc and is_perspective arguments in camera module

Many thanks to this release's contributors!
@WYJSJTU @juxuan27 @mingyuan-zhang @ttxskk @YongtaoGe @LazyBusyYang @pangyyyyy

MMHuman3D Release v0.6.0

Highlights

• Add HumanDataCache that requires 96% less RAM during training
• Refactor differentiable renderers and support UV map rendering
• Support slice/concat operations for HumanData

New Features

• Add HumanDataCache to address OOM issue when training using HumanData (e.g., HMR training requires 96% less RAM)
• Refactor differentiable renderers and support UV map rendering
• Support slice/concat operations for HumanData
• Upgrade SMPLify to support more losses, output HumanData, and more
• Refactor keypoint mapping and support mask-free conversion

Bug Fixes

• Fix critical bugs in data converters (for SURREAL, AGORA and more)
• Fix vertices-related bug in visualization toolbox
• Fix shape and device in tensor2dict
• Fix small bugs in HybrIK
• Fix OOM error for demo
• Fix doc format issue
• Remove duplicate limbs in HumanData Update dependency (e.g. cdflib, ffmpeg)

MMHuman3D Release v0.5.0

Highlights

• Support new data structure SMC for new dataset HuMMan that will soon be released
• Support for multi-GPU training/testing without slurm
• Support training-time validation and additional metrics such as PVE
• Bug fixes in data augmentation for more stable training
• Stronger HybrIK baseline (PA-MPJPE 49.02 on 3DPW)

New Datasets

• Add data converter for HuMMan dataset
• Add support for training with HuMMan dataset

New Features

• Add a new data structure SMC with SMCReader
• Add support for multi-GPU training/testing without slurm
• Add eval hook that supports validation during training and additional metrics such as PVE
• Support rigid transformation of SMPL parameters
• Add video-based inference pipeline to support VIBE demo
• CameraParameter accepts numpy and torch tensor and K, R, T can be obtained by a single method

Better Methods

• Reproduce a stronger HybrIK baseline (PA-MPJPE 49.02 on 3DPW)

Bug Fixes

• Fix a bug in data augmentation that causes keypoint2d loss not converging
• Fix a bug in SMPL rotation augmentation
• Fix bugs in data converters
• Fix GPU memory wastage due to unnecessary initialization

Documentation

• Use shared menu from OpenMMLab theme
• Add metafile

MMHuman3D Release v0.4.0

Refactoring

• Registration-based methods are moved to models/, under the new module abstraction named “registrants”
• Body model wrappers have their own module abstraction named “body_models”

Datasets

• Add support for GTA-Human dataset
• Add support for datasets needed by VIBE

New features

• Upgrade SMPLify: batch size adaptation and use standard camera module
• Upgrade camera: helper functions such as concat, parameter conversion and value type check
• Upgrade renderer: refactored to use registry, refactored smpl visualization

Documents

• Update README, tutorials, and readthedocs

Bug fixes:

• Unit tests: test_cache that raises time comparison error
• Visualization: holes in rendered mesh
• Data converters: image paths, and H36M and LSP

CICD

• Deploy workflow on GitHub
• Add codedev reports in workflow

MMHuman3D Release v0.3.0

Main Features

• Supports registration-based methods: SMPLify and SMPLify-X
• Supports regression-based methods: HMR, SPIN, VIBE, and HybrIK
• Supports 16 datasets with the unified data format HumanData
• Supports differentiable visualization of parametric models

The actual commit id for this tag is 8255b06