Open XinGP opened 3 months ago
Hi, As described in the paper, we use ensemble techniques to get the test set results on the Argoverse 1, which is a quite standard procedure. We train several models with different random seeds. For the focal agent in a sequence, we gather the results produced by all these models, and use k-means clustering to find K=6 distinct categories and their corresponding weight. The final trajectories are calculated using the average of trajectories in each category, and the probability score is the mixture weight from k-means clustering.
Hello author! May I ask how the test set results obtained from model ensemble were obtained? Is it through the ensemble of several models obtained from a single training session or through multiple training sessions? Can you disclose the specific ensemble details?