search

Daniil-Osokin / lightweight-human-pose-estimation.pytorch

Fast and accurate human pose estimation in PyTorch. Contains implementation of "Real-time 2D Multi-Person Pose Estimation on CPU: Lightweight OpenPose" paper.
Apache License 2.0
2.08k stars 473 forks source link

Could ~40 AP be used in a real-world scenario？ #105

Closed sunbin1357 closed 4 years ago

sunbin1357 commented 4 years ago

Could ~40 AP be used in a real-world scenario？Have you made any recent progress in the field of pose estimation？

Daniil-Osokin commented 4 years ago

Hey, @sunbin1357! Yes, it can be used for real-world scenario. Off course, it totally depends on the use case, but I know, it was successfully used for action recognition and for pose tracking in sport. It worth to try. The focus of this work is on real-time inference for pose estimation. The quality is ok, but problems may arrive for small sized poses. For this case we developed fast and accurate (~70% AP) top-down method in neighbor repository, but it requires to run person detector first. Moreover, I also have the plans to improve this work in terms of inference speed with similar quality, however it needs some free time and effort :).

sunbin1357 commented 4 years ago

your paper report the AP of openpose with 5 stages is 48.6, however the result in other papers is 60+.

Daniil-Osokin notifications@github.com 于2020年8月21日周五 上午4:42写道：

Hey, @sunbin1357 https://github.com/sunbin1357! Yes, it can be used for real-world scenario. Off course, it totally depends on the use case, but I know, it was successfully used for action recognition and for pose tracking in sport. It worth to try. The focus of this work is on real-time inference for pose estimation. The quality is ok, but problems may arrive for small sized poses. For this case we developed fast and accurate (~70% AP) top-down method in neighbor repository https://github.com/Daniil-Osokin/gccpm-look-into-person-cvpr19.pytorch, but it requires to run person detector first. Moreover, I also have the plans to improve this work in terms of inference speed with similar quality, however it needs some free time and effort :).

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/Daniil-Osokin/lightweight-human-pose-estimation.pytorch/issues/105#issuecomment-677893230, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALNJVR6JUN5CZ44MUZCHS6TSBWDDPANCNFSM4QFZMGIQ .

Daniil-Osokin commented 4 years ago

Yeah, it is a tricky part. 60+% AP is achieved by multiscale inference (4 scales), then each found pose is additionally refined with CPM model, it is written in original OpenPose paper.

sunbin1357 commented 4 years ago

Do you have a demo that applies to action recognition or other real-world scenario

Daniil-Osokin commented 4 years ago

This repository contains just the pose estimation.

Daniil-Osokin commented 4 years ago

Hope, it is clear now.