Abstract
Human motion prediction model has applications in various fields of computervision. Without taking into account the inherent stochasticity in theprediction of future pose dynamics, such methods often converges to adeterministic undesired mean of multiple probable outcomes. Devoid of this, wepropose a novel probabilistic generative approach called Bidirectional Humanmotion prediction GAN, or BiHMP-GAN. To be able to generate multiple probablehuman-pose sequences, conditioned on a given starting sequence, we introduce arandom extrinsic factor r, drawn from a predefined prior distribution.Furthermore, to enforce a direct content loss on the predicted motion sequenceand also to avoid mode-collapse, a novel bidirectional framework isincorporated by modifying the usual discriminator architecture. Thediscriminator is trained also to regress this extrinsic factor r, which is usedalongside with the intrinsic factor (encoded starting pose sequence) togenerate a particular pose sequence. To further regularize the training, weintroduce a novel recursive prediction strategy. In spite of being in aprobabilistic framework, the enhanced discriminator architecture allowspredictions of an intermediate part of pose sequence to be used as aconditioning for prediction of the latter part of the sequence. Thebidirectional setup also provides a new direction to evaluate the predictionquality against a given test sequence. For a fair assessment of BiHMP-GAN, wereport performance of the generated motion sequence using (i) a critic modeltrained to discriminate between real and fake motion sequence, and (ii) anaction classifier trained on real human motion dynamics. Outcomes of bothqualitative and quantitative evaluations, on the probabilistic generations ofthe model, demonstrate the superiority of BiHMP-GAN over previously availablemethods.