Beyond Perfect Scores: Leakage-Aware Multi-Corpus Evaluation of Low-Resolution ToF and Depth Gesture Recognition

Touchless interfaces for embedded systems require models that are both compact and evaluated under protocols that reflect realistic generalization. Public low-resolution gesture datasets can obscure this requirement because repeated samples, temporal neighbors, or subject identity may leak across train and test partitions. This study presents a leakage-aware multi-corpus benchmark in which four public corpora are evaluated within their own dataset-specific partitions after conversion to a common 8 × 8 depth representation when needed: a public IEEE DataPort ToF CSV with 8400 frames from four gestures, DS.GestureRecognition_TOF.1.0 with 3109 frames extracted from 40 low-resolution ToF sequences and three gestures, the depth component of the public Kinect+Leap dataset with 1400 samples from 14 subjects and 10 gesture classes, and the public Senz3D depth archive with 1320 samples from 4 subjects and 11 gesture classes. No blurred RGB image dataset, deblurring target, or synthetic blur augmentation is produced in the experiments; the only spatial reduction applied to external depth corpora is deterministic block averaging for a shared low-resolution depth input. The term multi-corpus is used here to clarify scope: the experiments do not train on one dataset and test on another, but instead compare optimistic random splits against structure-aware alternatives matched to each corpus: chronological block-aware, sequence-aware, and subject-wise evaluation. On the original IEEE ToF benchmark, the best macro F1-score decreases from 1.0000 under random splitting to 0.8494 under block-aware evaluation. On DS.GestureRecognition_TOF.1.0, random splitting again reaches 1.0000, whereas the best sequence-aware result falls to 0.6697. On the downsampled Kinect+Leap depth benchmark, random frame classification reaches 0.7639 macro F1, while subject-wise evaluation drops to 0.0566. On Senz3D reduced to 8 × 8, random frame classification reaches 0.9068 macro F1, while subject-wise evaluation drops to 0.0717. Device-side validation on an NVIDIA Jetson Nano Developer Kit preserved a macro F1-score of 0.8667 for the selected GRU on the primary IEEE ToF dataset with 14.48 ms/window steady-state CPU latency. The results show that low-resolution gesture recognition conclusions are strongly protocol- and corpus-dependent: compact temporal models remain viable for genuine ToF streams, but naive random splitting can substantially overstate generalization, and subject-wise recognition after aggressive 8 × 8 conversion remains difficult for the evaluated datasets and architectures.