You propose to use SVD to construct ET space for normalized trajectory A(L*N). I'm not quite sure if N here represents the number of all tracks in each training set ("eth" "hotel" "univ" "zara1" "zara2")?
When I was downloading the pretrained model file, I found that different datasets have a corresponding "static.dist" coefficient, such as the ETH dataset's "static.dist": 0.419, the ZARA1 dataset's "static.dist": 0.338, and so on. Now I want to use the SDD dataset for training. How should I determine the corresponding "static_dist"? How much is it confirmed?
The method has some anchors of the coefficients. You use the minimal difference between the reconstruction and ground-truth trajectory to choose the right anchor when training. However, during inference, since we don't have the GT trajectory, how to choose the right anchor and do reconstrution?