search

exitudio / GaitMixer

Official repository for "GaitMixer: Skeleton-based Gait Representation Learning via Wide-spectrum Multi-axial Mixer"
24 stars 1 forks source link

The difference in accuracy of reproduction. #6

Open ZziTaiLeo opened 11 months ago

ZziTaiLeo commented 11 months ago

image Hi， I just download the data.zip from gaitgraph for casia-b_poses_train_valid.csv and casia-b_valid.csv, then use experiments/train.sh to train model, but the accuracy looks have a bit difference from paper. I have not adjusted any hyperparameters or network structure. How can I adjust it to reproduce this accuracy. Looking forward to your early reply! Tks！

exitudio commented 11 months ago

Hi, Thank you for being interested in our work. the variance we're seeing is likely because the CASIA-B dataset is relatively small, and the inclusion of bag (BG) and coat (CL) data adds some noise. We ran the experiment 8 times and reported the best result. Note that, this pattern is not unique to our model—other methods, like GaitGraph, exhibit similar behavior.

ZziTaiLeo commented 11 months ago

Thank you for your prompt reply! I only saw it now. Although you said you had the best results in eight experiments, I still cannot accept the current results and cannot find a corresponding solution to solve it. Are you downloading the processed skeleton dataset directly?

exitudio commented 11 months ago

Yes. We use the preprocessed data from GaitGraph.