Edges Are All You Need: Robust Gait Recognition via Label-Free Structure

A research team has published a paper titled 'Edges Are All You Need: Robust Gait Recognition via Label-Free Structure,' introducing a breakthrough framework called SketchGait. The work, led by Chao Zhang and colleagues, challenges the dominant paradigms in gait recognition, which rely on either sparse human silhouettes or parsing-based methods that require heavily labeled data and upstream human parsers. Instead, the researchers propose using a novel 'sketch' visual modality. This modality extracts dense, part-level structural cues—like limb articulations and occlusion contours—directly from RGB images using standard edge detectors, all without needing explicit semantic labels.

The core innovation is the identification of an underexplored design space: using dense structural information without semantic supervision. SketchGait is a hierarchically disentangled multi-modal framework with two independent streams. One stream learns from the new label-free sketch data, while the other can utilize traditional parsing data, with a lightweight fusion branch to combine their complementary strengths. This architecture allows the model to capture high-frequency details that silhouettes miss and avoids the instability of parser-dependent methods.

Extensive validation on the SUSTech1K and CCPG datasets demonstrates the framework's superiority. SketchGait achieved a 92.9% Rank-1 accuracy on SUSTech1K and a 93.1% mean Rank-1 on CCPG, setting a new benchmark. The paper argues that the label-free sketch and label-guided parsing are semantically decoupled but structurally complementary, making their combined use particularly powerful for a non-intrusive biometric technique crucial for security applications.