Name
simple_pv_lstm
Description
A simple RNN-based model to learn and predict the future pose and trajectory. we used LSTM blocks to train the model. One of the main challenges of training this model was the variety of metrices which were not necessarily in the same line. For instance, optimizing the L1 loss did not necessarily result in better ADE score, and on top of that, VIM and VAM and the visibility masks made the problem more complex. On top of that, model could achieve good performance on training dataset but usually it resulted to much worse performance on validation dataset.
Publication title
null
Publication authors
null
Publication venue and year
null
Publication URL
Language(s)
Hardware
Website
Source code URL
Creation date
2021-11-22 16:36:15
| Metric | 80/100 ms | 160/240 ms | 320/500 ms | 400/640 ms | 560/900 ms |
|---|---|---|---|---|---|
| PoseTrack VAM | 25.364754275861962 | 43.85653479736883 | 73.62615195392183 | 85.69684350224928 | 101.46226794141538 |
| PoseTrack VIM | 12.074983204446129 | 22.332417011417842 | 41.50673671331702 | 50.66875901813145 | 64.783411436018 |
| 3DPW VIM | 23.425032651592957 | 42.56655433585515 | 78.55781687760147 | 97.91094035501958 | 137.27033711155144 |