Roadmap of MMPose 1.x

[Tau-J]

Development Plan 2023

We are excited to announce the release of MMPose 1.0.0 as a part of the OpenMMLab 2.0 project!

MMPose v1.0.0 is a major update, including many API and config file changes, and most of the code (and config files) in MMPose 1.0 will not be compatible with 0.x version.

Here are the version correspondences between mmdet, mmcv and mmpose:

• MMDetection 2.x <=> MMPose 0.x <=> MMCV 1.x
• MMDetection 3.x <=> MMPose 1.x <=> MMCV 2.x

Notes: mmcv-full is only for mmcv 1.x, so you have to uninstall it first, then install mmcv 2.x.

Migration

Currently, a part of the algorithms have been migrated to v1.0.0, and the remaining algorithms will be completed in subsequent versions. We will show the migration progress in the following list.

Algorithm

Algorithm	Status
MTUT (CVPR 2019)
MSPN (ArXiv 2019)	done
InterNet (ECCV 2020)	in progress
DEKR (CVPR 2021)	done
HigherHRNet (CVPR 2020)
DeepPose (CVPR 2014)	done
RLE (ICCV 2021)	done
SoftWingloss (TIP 2021)	done
VideoPose3D (CVPR 2019)	done
Hourglass (ECCV 2016)	done
LiteHRNet (CVPR 2021)	done
AdaptiveWingloss (ICCV 2019)	done
SimpleBaseline2D (ECCV 2018)	done
PoseWarper (NeurIPS 2019)
SimpleBaseline3D (ICCV 2017)	done
HMR (CVPR 2018)
UDP (CVPR 2020)	done
VIPNAS (CVPR 2021)	done
Wingloss (CVPR 2018)	done
DarkPose (CVPR 2020)	done
Associative Embedding (NIPS 2017)	in progress
VoxelPose (ECCV 2020)
RSN (ECCV 2020)	done
CID (CVPR 2022)	done
CPM (CVPR 2016)	done
HRNet (CVPR 2019)	done
HRNetv2 (TPAMI 2019)	done
SCNet (CVPR 2020)	done

• If your algorithm has not been migrated, you can continue to use the 0.x branch and old documentation.
• If you wish to help us to migrate, please contact us. We will much appreciate your contribution!

Datasets

Dataset	Status
COCO	done
MPII	done
MPII-TRB	done
AI Challenger	done
CrowdPose	done
OCHuman	done
MHP	done
PoseTrack18	done
sub-JHMDB	done
COCO-WholeBody	done
Halpe	done
300W	done
WFLW	done
AFLW	done
COFW	done
COCO-WholeBody-Face	done
OneHand10K	done
FreiHand	done
CMU Panoptic HandDB	done
InterHand2.6M	in progress
RHD	done
COCO-WholeBody-Hand	done
DeepFashion	done
DeepFashion2	done
Animal-Pose	done
AP-10K	done
Horse-10	done
MacaquePose	done
Vinegar Fly	done
Desert Locust	done
Grévy’s Zebra	done
ATRW	done
Human3.6M	done
CMU Panoptic	in progress
Campus/Shelf	in progress

New Algorithms

• ViTPose (1.x in [Feature] Support ViTPose #1876) (0.x in [Feature] Support ViTPose #1937)
• RTMPose [projects] in [Feature] Add RTMPose #1971
• YOLO-Pose [projects] in [Feature] Support YOLO-Pose #2020
• PETR (in progress)
• MotionBERT
• Uniformer [projects] in [Feature][MMSIG] Add UniFormer Pose Estimation to Projects folder #2501

New Datasets

• DeepFashion2 in [Feature] Add support for deepfashion2 dataset  #2201
• Human-Art [github] in [Feature] Support Human-Art Dataset #2304
• LaPa [github] in [Feature] Support LaPa Dataset #2281
• AnimalKingdom [github] in [Feature] Support AnimalKingdom dataset #2139
• UBody (in progress) [github] in [Feature] Support ubody dataset (2d keypoints) #2588

[Ben-Louis]

How about supporting UniFormer?

[tojimahammatov]

Hi, is there any news on migrating Associative Embedding algorithm? It would be very great to see it in 1.x version of mmpose. When are you planning to publish it?

It's a very important algorithm in most of the models and I'm really looking forward to having it. Or can you hint some ways to implement? I just need it for one of my ongoing researches. Thank you.

[tojimahammatov]

@Tau-J hello, sorry for the ping, is there any update on associative embedding? Could you please answer my questions above? I was working on an important task that is related to this algorithm. Thank you.

[Tau-J]

@Tau-J hello, sorry for the ping, is there any update on associative embedding? Could you please answer my questions above? I was working on an important task that is related to launch some services to a production. Thank you.

Sorry for the late reply. The algorithm of 'Associate Embedding' is still in the process of migration. As the community contributor responsible for this task is not a full-time developer, we cannot guarantee a specific completion timeline. If you are also willing to contribute to the migration, we would greatly appreciate it. You can follow our 1.x documentation to understand the new version architecture and proceed with migrating the algorithm. If you meet any issue during migration, please feel free to contact us.

[tojimahammatov]

@Tau-J hello, sorry for the ping, is there any update on associative embedding? Could you please answer my questions above? I was working on an important task that is related to launch some services to a production. Thank you.

Sorry for the late reply. The algorithm of 'Associate Embedding' is still in the process of migration. As the community contributors responsible for this task is not a full-time developer, we cannot guarantee a specific completion timeline. If you are also willing to contribute to the migration, we would greatly appreciate it. You can follow our 1.x documentation to understand the new version architecture and proceed with migrating the algorithm. If you meet any issue during migration, please feel free to contact us.

Never mind, thanks for the reply. I got it, I will try to migrate and contribute then. I do contact if I need some help or issue. Thank you.

[Omfff]

When will motionbert be merged into the main branch? Currently, I only see it in the dev branch.

[Ben-Louis]

When will motionbert be merged into the main branch? Currently, I only see it in the dev branch.

Yes. The dev-1.x branch will be merged into the main branch recently

[thehiddenwaffle]

It would be really wonderful for RTM's wholebody applications if pose lifting was available using something like one of the models referenced in H3WB, though I'm inexperienced and don't know if this is viable.

[zbl929]

Can you migrate the top_down_pose_tracking_demo_with_mmdet.py from mmpose 0.x to mmpose 1.x, thank you !

[gdtiti]

请问支持unity插件的事情有开始实施吗？

[tojimahammatov]

@Tau-J @Ben-Louis hi guys, sorry for the ping. We've been working with pose estimation task for some time. Until now, we could achieve some feasible results by successfully adapting pose estimation models to our domain space. However, it seems that there are still things to improve. A few issues like track fragmentation (wrong track is found for short time), and detecting tracks with black dress/shirts etc. can be observed from time to time.
Can I ask you to make a contact or a call for at least consulting for us? Our company representatives and tech team will be very glad to listen to your invaluable feedbacks and ideas. I know that you also work for full-time and you must be too busy. But, I believe that it would cause a great progress for our development. My email is tojimahammatov@gmail.com. We look forward to hearing from you soon. Thanks.