EVENT DOSSIER
To improve the performance of single-eye depth estimation in complex driving scenarios and the efficiency of deployment on edge devices, researchers proposed the FlexDepth model family. This model adopts a two-stage static-dynamic decoupling training strategy and introduces a scale-driven decoder (SDD) that can dynamically select components based on scale size to achieve efficient feature fusion. In standard driving benchmarks, FlexDepth achieves advanced performance across all scales without the need for auxiliary information, with extremely low computational overhead. Its smallest model, Flex-Nano, requires only 0.7 GFLOPs, achieving 37.6 FPS on mobile platforms, and possesses excellent zero-sample generalization capabilities and reliable real-time perception performance.
arXiv:2607.00736v3 FlexDepth 提出自监督单目深度估计模型
为应对现有单目深度估计模型在复杂驾驶环境中性能下降及难以部署于边缘设备的问题，研究者提出 FlexDepth 这一面向挑战道路场景的尺度驱动自监督单目深度估计模型家族。该模型采用两阶段静态 - 动态解耦训练策略，并引入尺度驱动解码器（SDD）以根据尺度大小动态选择组件，实现高效特征融合与高精度深度图输出。在标准驾驶基准测试中，FlexDepth 无需辅助信息即可在任意尺度下达到最先进性能且计算开销极低；其最小模型 Flex-Nano 仅需 0.7 GFLOPs，在移动平台上实现 37.6 FPS，确保可靠的实时感知并具备优秀的零样本泛化能力。