A 4D point-cloud model can read motion or shape: freezing one arm's movement and stripping its shape splits the answer
Synopsis
Before deployment, a 4D point-cloud action model cannot say whether it relies on how a body part moves or on what shape it has. Kinematic Causal Attribution queries only model outputs and splits each spacetime region's importance into a motion term and a geometry term using two interventions: freezing motion and removing shape. On three official backbones trained on MSR-Action3D, the same arm's motion share is 0.83 versus 0.60, 0.86 versus 0.30 and 0.92 versus 0.61 across motion and shape tasks, while a model-blind motion prior marks the arm identically in both.
Interpretation
Explanations of 4D point-cloud video models can now answer a question no prior method could pose: whether a decision depends on how a structure moves or on the shape it has in each frame. Earlier 3D attribution is single-frame, video attribution is confined to RGB pixels, gradient maps need white-box access, attention read-outs do not exist on attention-free state-space backbones, and point-level perturbations assume cross-frame point correspondence that depth and LiDAR data lack because the surface is re-sampled every frame. On three official backbones trained on MSR-Action3D, two binary tasks built from real clips differ only in the motion or only in the shape of the same right arm, and the motion share of the arm's importance follows the defining cue: 0.83 versus 0.60 on P4Transformer, 0.86 versus 0.30 on PSTNet, 0.92 versus 0.61 on Mamba4D, while a model-blind motion prior marks the arm identically in both tasks (EBPG 0.28 versus 0.28).
The split costs no faithfulness: the necessity map stays at or above the strongest baseline in deletion tests and loses its correlation with the trained model once the weights are randomised. On P4Transformer the context estimator's deletion AUC is 0.179, below region occlusion at 0.233 and Integrated Gradients at 0.282; it is 0.217 on PSTNet and 0.269 on Mamba4D, statistically tied with region occlusion there. Under cascading randomisation from the head down, the rank correlation of KCA's map with the trained model's drops to 0.12 and stays at 0.10 with every block randomised, while GradInput keeps a correlation of 0.81.
One code path explains convolutional, attention-based and attention-free state-space backbones using forward queries alone. Attention read-outs exist only for models such as P4Transformer that have attention, and Mamba4D has none to read. On one A100 GPU the local estimator takes 0.5 s per clip and the context estimator 2.3 s, comparable to OccAM at 1.5 s and Integrated Gradients at 1.9 s, with no backward pass.
Perspective
The demonstrated setting is action recognition on MSR-Action3D: three official backbones trained from scratch on centred 24-frame clips of 2048 points per frame, with ground-truth metrics covering the 275 benchmark clips that have a complete skeleton. The method targets practitioners auditing 4D point-cloud video models before deployment and queries only model outputs, requiring no point correspondences.
No pretrained weights exist, so all three backbones are trained from scratch, and PSTNet and Mamba4D fall short of their published accuracies; at full scale the P4Transformer motion-task model leaves chance in only 3 of 6 training seeds. Skeleton ground truth is approximate and marks where motion is, not where the model should look. Regions come from a k-means tracker that can drift under fast or occluded motion, and point-level maps are less stable under sensor re-sampling than gradient maps.