Sketch It Out: Exploring Label-Free Structural Cues for Multimodal Gait Recognition

Gait recognition is a non-intrusive biometric technique for security applications. Existing studies mainly rely on silhouette, skeleton map and parsing representations. Although silhouette captures global body shape, it is structurally sparse and lacks internal details. Skeleton maps encode different body parts but still miss fine shape cues. Parsing provides richer part-level structures, yet its effectiveness depends heavily on the quality of upstream parsing, such as label granularity and boundary precision. These distinct characteristics suggest the need for a more unified view on gait visual representations. To this end, we revisit existing representations from the perspectives of structural edge density and semantic richness. Silhouette-based methods rely on sparse boundary structures with single semantic foreground. Skeleton map-based methods are also structurally sparse, but encode different body parts with richer semantic information. In contrast, parsing-based methods use denser structural cues with finer semantic labels. This view reveals an underexplored direction: learning dense part-level structural information without explicit semantic labels. To fill this gap, we introduce Sketch as a new visual modality for gait recognition. Sketch is a representation-level label-free modality that extracts high-frequency structural cues from foreground-masked RGB without encoding explicit semantic part labels. Our experiments evaluate the effectiveness and limitations of Sketch under diverse conditions. The results further demonstrate that Sketch and Parsing effectively complement each other: Sketch compensates for structural boundaries missing in Parsing, while Parsing mitigates texture noise and other irrelevant patterns over-extracted by edge-based Sketch. To better exploit the complementary characteristics of these two modalities, we propose SketchGait, a multimodal framework with two independent streams for modality-specific learning and a lightweight early-fusion branch. Furthermore, by incorporating the Silhouette modality, we extend it to SketchGait++, which further improves the overall performance. Extensive experiments on SUSTech1K and CCPG demonstrate the effectiveness of the proposed modality and framework.