Work

[Re] From goals, waypoints and paths to longterm human trajectory forecasting

Year: 2022
Type: article
Abstract: Scope of ReproducibilityThe following paper is a reproducibility report for From Goals, Waypoints & Paths To Long Term Human Trajectory Forecasting [1].The basic code was made available by the author at this https url.We have verified all claims and results from the experiments mentioned in the paper to support the claims.The central claim of YNet is that it sets state-of-the-art short and long-term prediction standards by a multi-modal network employing both segmentation matrices and past trajectory heat-maps together. MethodologyThe model essentially combines a segmentation map and past trajectory heatmaps to encode a combined input to three sub-networks modelled after the U-Net architecture [2].The author's code was used to benchmark the claims, and some experiments were performed thereafter.Free-to-use platforms like Google Colaboratory and Kaggle were used to train these models.We have reproduced the code base in PyTorch Lightning (originally in PyTorch Ignite) and found consistent results across the board. ResultsThrough our testing, we were able to come within 2% of the proposed metrics on certain datasets like Stanford Drone Dataset (SDD) and ETH/UCY, implying the author's claims are sanguine and reproducible on varied hardware.However, certain results such as long-term predictions on the Intersection Drone (InD) dataset were quite different; the probable reasons for which have been discussed. What was easyObtaining the proposed results on the SDD and InD datasets was easy.Well-documented interactive notebooks for training and testing along with the requisite data for SDD were provided with the codebase.The code could be run with minimal changes overall.
Language: English
Open Access status: green
OpenAlex ID: W4281399345
FWCI:
Cites: 14
Cited by:
Related to: 10