Multiple Object Forecasting: Predicting Future Object Locations in Diverse Environments

SimAug: Learning Robust Representations from Simulation for Trajectory Prediction

“…Notably, the Stanford Drone Dataset (SDD) [50] is used in many works [53,13,30] for trajectory prediction with drone videos. Other works have also looked into pedestrian prediction in dashcam videos [42,57,25,28] and first-person videos [70,58]. Many vehicle trajectory datasets [6,8,71] have been proposed as a result of self-driving's surging popularity.…”

“…A notable bottleneck for existing works is that the current model is closely coupled with the video cameras on which it is trained, and generalizes poorly on new cameras with novel views or scenes. For example, prior works have proposed various models to forecast a pedestrian's trajectories in video cameras of different types such as stationary outdoor cameras [44,32,1,19,29,37], drone cameras [53,13,30], ground-level egocentric cameras [70,47,58], or dash cameras [42,57,8]. Models that rely on RGB pixel inputs are especially vulnerable to view changes [34,33,53].…”

Multimodal Future Localization and Emergence Prediction for Objects in Egocentric View with a Reachability Prior

“…TraPHic [10] exploits the interaction between nearby heterogeneous objects. DTP [49] and STED [48] use encoder-decoder schemes using optical flow and past locations and scales of the objects. Yao et al [63] added the planned egomotion to further improve the prediction.…”

“…STED [48]. STED is a spatial-temporal encoderdecoder that models visual features by optical flow and temporal features by the past bounding boxes through GRU encoders.…”

“…All these baselines predict the future trajectory of either pedestrians [49,48,7] or vehicles [63]. Thus, we re-trained them on nuScenes [9] to handle both pedestrian and vehicle classes.…”