search

cure-lab / DeciWatch

[ECCV 2022] Official implementation of the paper "DeciWatch: A Simple Baseline for 10x Efficient 2D and 3D Pose Estimation"
Apache License 2.0
177 stars 15 forks source link

smoothnet & deciwatch, which is better? #1

Closed ykk648 closed 2 years ago

ykk648 commented 2 years ago

or we can use both?

ailingzengzzz commented 2 years ago

Hi @ykk648,

They have different focuses. For SmoothNet, it is used for smoothing jitters to get a smooth curve or pose sequence, which generalizes well across different 2D,3D, or 6D modalities (share weights across spatial dimensions). It needs to input the whole detected poses.

For DeciWatch, it focuses on high efficiency with lower mean Flops and faster inference speeds via skipping and estimating highly sparse input frames, e.g., estimating one frame in every ten frames. It inputs sparse detected poses. Because DeciWatch utilizes the continuity of human motions, it can also output a smooth pose sequence after RecoverNet.

Therefore, you can choose which method to use depending on your needs.

ykk648 commented 2 years ago

@ailingzengzzz Thanks for your response, I have tested DeciWatch under mmhuman3d framework , the results looks similar to slerp for me, it can replace filter to some extent , SmoothNet looks like a CNN-version filter with motion supervision, I'm willing to test it after next update of mmhuman3d.

ailingzengzzz commented 2 years ago

@ykk648 The learning-based methods will show their effectiveness when the inputs are quite noisy. If the inputs look good, traditional filters or interpolation manners will also work well.