Computer Vision News - April 2019

Results of temporal regularization on object tracking estimates - errors in translation as [mm] and rotation in [degrees]: Results on camera pose estimates provided by PoseNet on the Cambridge Landmarks dataset. Values are in [m] of translation and in [degrees] for rotation: See the full article for more detailed camera tracking and object tracking results. Conclusion The paper presented long short-term memory Kalman filter (LSTM-KF) model, a tool that makes motion modeling and estimation problems much easier, and makes it possible to learn rich models from difficult-to-model real-world data. Utilizing a wide array of experiments, the authors showed LSTM-KF outperforms the standard Kalman filter and standard LSTM, achieving state of the art performance on 3 widely divergent tasks -- just to illustrate, their model reduced the error in localization of body joints on the Human3.6M database by 13.8%, from 82.3 mm to 71.0 mm. Research 10 Research Computer Vision News …a tool that makes motion modeling and estimation problems much easier, and makes it possible to learn rich models from difficult-to-model real-world data.