Human motion prediction trains in stages: CPC cuts PGBIG's 3DPW average error from 53.8 to 44.7 mm
Synopsis
Human motion prediction models usually train near-future and far-future poses in one stage, and the high-uncertainty far-future fit suppresses short-term learning. CPC and CPC++ split the future sequence into segments, add farther prediction tasks stage by stage, and use learnable prior compensation factors to estimate early knowledge lost during multi-stage training. With PGBIG as backbone, CPC lowers average MPJPE on Human3.6M, CMU-MoCap and 3DPW to 52.9, 35.4 and 44.7 mm, against baselines of 54.4, 36.4 and 53.8 mm.
Interpretation
Human motion prediction training is split into stages, so the model first learns near-future poses and then gradually takes on farther future segments. Earlier approaches treated predictions at different moments as one task and learned short-term and long-term prediction together in a single stage. On the Human3.6M, CMU-MoCap and 3DPW benchmarks with PGBIG as backbone, CPC lowers average MPJPE from the baseline 54.4, 36.4 and 53.8 to 52.9, 35.4 and 44.7 mm, and CPC++ further lowers it to 52.7 mm on Human3.6M.
Knowledge stored in model parameters after each stage is gradually overwritten in later stages, and the compensation factor keeps an explicit account of that loss. CPC summarizes prior forgetting across all earlier subtasks with one learnable random variable, PCF, while CPC++ uses a fine-grained prior compensation factor, FGPCF, that estimates forgetting for each subtask separately. FGPCF comes from a prior loss estimation module that first lets multi-head self-attention exchange features across time steps of all tasks, then applies per-task max-pooling and an MLP with sigmoid mapping to output a forgetting estimate between 0 and 1.
The compensation factor is trained jointly with model parameters by minimizing an upper bound of the objective rather than the objective itself. Determining the compensation factor's value directly is difficult, and the derived upper bound turns it into a coefficient that controls each task's loss weight. On Human3.6M with PGBIG as backbone, task 1 prediction error rises by 0.83 without PCF, by only 0.27 with PCF, and by only 0.14 with FGPCF.
Perspective
The demonstrated scope is three 3D human pose benchmarks, Human3.6M, CMU-MoCap and 3DPW, with LTD, MotionMixer, siMLPe and PGBIG as backbones, 10 input frames and 25 or 30 predicted frames, evaluated by MPJPE. The method is used only in training, and inference matches the baseline model. The authors list extension to more general prediction tasks and backbone models as future work.
The derivation of the compensation-factor upper bound relies on a confidence assumption; the authors count 97.8% of samples on Human3.6M satisfying it, and for the rest the gap between objective and upper bound stays in a reasonable range. On 3DPW, CPC++ performs comparably to CPC, which the authors attribute to the smaller dataset and more complex motions making the finer FGPCF harder to learn for a model with more parameters. Average error stays roughly flat as the task count grows from 3 to 5 or 8.
