Note
All these settings could be changed before making (transforming) dataset files in the corresponding py files in the codes folder.
ETH-UCY and Stanford Drone Dataset
When making dataset files, we process the ETH-UCY dataset files by the leave-one-out strategy.
For the SDD, we split video clips with the dataset split method from SimAug (36 train sets + 12 test sets + 12 val sets).
Especially, following other researchers, the univ13 split is not split exactly according to the literal leave-one-out approach.
The univ13 split (ETH-UCY) takes univ and univ3 as test sets, and other sets {eth, hotel, unive, zara1, zara2, zara3} as training sets.
Differently, the univ split only includes univ for testing models.
Note
Please note that we do not use dataset split files like TrajNet on ETH-UCY and SDD for several reasons.
For example, the frame rate problem in ETH-eth sub-dataset, and some of these splits only consider the pedestrians in the SDD dataset.
We process the original full-dataset files from these datasets with observations = 3.2 seconds (or 8 frames) and predictions = 4.8 seconds (or 12 frames) to train and test the model.
See details in this issue.
NBA SportVU
Following most previous works, we only sample about 50K samples on this dataset.
Thus, the corresponding split name has become nba50k.
Among the 50K samples,
65% Train sets: About 32.5K training samples;25% Test sets: About 12.5K test samples;10% Val sets: About 0.5K samples.
nuScenes
Since the official test set in the nuScenes dataset has no labels, we use the following method to split the training set:
550 Train sets: Randomly sampling 550 video clips from the official 850 training sets;150 Val sets: Randomly sampling the other 150 video clips from the official 850 training sets;150 Test sets: We treat the official 150 val sets as our test sets.
Human3.6M
Following previous settings, we use all data from subjects [1, 6, 7, 8, 9] to train the model, subjects [11] to validate, and subjects [5] to test.
When making the training samples, we sample observations with the frequency of 25Hz (i.e., the sample interval is 40ms) and use t_h = 10 frames (400ms) of observations from all subjects to predict their possible trajectories for the next t_f = 10 frames (400ms).