Computer Vision News - April 2019

and 5,000 frames. For this paper, the authors use for training the sets of movements marked as S1, S5, S6, S7 and S8 and for testing the sets S9 and S11. For more details about the those sets, see the Human3.6M website . The table below provides detailed quantitative results: Below follow the qualitative results for the evaluation on the Human3.6M database. The ground truth is in red and the model’s estimate is in green. Based on the initial CNN estimate, we compare outputs of standard Kalman, standard LSTM and the authors’ LSTM-KF model. LSTM-KF shows significant improvement over the other methods, especially for the arm joints and leg joints. 9 Research Computer Vision News Long Short-Term Memory Kalman Filters