Progressive Gradient-Guided Self-Distillation Keypoint Detection

With the increasing popularity of autonomous driving and 3D reconstruction, keypoint detection, as a key link in visual localization, has become a hot topic in current research. However, existing keypoint detection methods rarely pay attention to the difficulty differences of samples and lack a progressive learning mechanism, which often leads to overfitting for simple samples and underfitting for complex samples, limiting the overall performance of the model. To address these issues, we propose a novel progressive gradient-guided self-distillation method (PG<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">${^{2}}$</tex-math></inline-formula>SD) for keypoint detection, which possesses self-evolutionary learning capabilities. Specifically, we propose a progressive gradient constraint strategy (PGCS) that dynamically adjusts the gradient contributions of different samples, enabling the model to adapt to the evolving learning capability during training. On this basis, we propose a gradient-guided self-distillation strategy (G<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">${^{2}}$</tex-math></inline-formula>SDS), which integrates seamlessly with PGCS to alleviate the insufficient feature representation of hard samples in the early training stage. We further design a novel loss function to achieve dynamic collaboration between PGCS and G<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">${^{2}}$</tex-math></inline-formula>SDS, allowing G<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">${^{2}}$</tex-math></inline-formula>SDS to adaptively adjust the self-distillation parameters through the PGCS. Experimental results on multiple benchmark datasets show that our method achieves state-of-the-art performance on image matching, visual localization, and 3D reconstruction tasks without designing a proprietary network, indicating broad application prospects.