Unsupervised Spike Depth Estimation via Cross-modality Cross-domain Knowledge Transfer

Neuromorphic spike data, an upcoming modality with high temporal resolution, has shown promising potential in real-world applications due to its inherent advantage to overcome high-velocity motion blur. However, training the spike depth estimation network holds significant challenges in two aspects: sparse spatial information for dense regression tasks, and difficulties in achieving paired depth labels for temporally intensive spike streams. In this paper, we thus propose a cross-modality cross-domain (BiCross) framework to realize unsupervised spike depth estimation with the help of open-source RGB data. It first transfers cross-modality knowledge from source RGB to mediates simulated source spike data, then realizes cross-domain learning from simulated source spike to target spike data. Specifically, Coarse-to-Fine Knowledge Distillation (CFKD) is introduced to transfer cross-modality knowledge in global and pixel-level in the source domain, which complements sparse spike features by sufficient semantic knowledge of image features. We then propose Uncertainty Guided Teacher-Student (UGTS) method to realize cross-domain learning on spike target domain, ensuring domain-invariant global and pixel-level knowledge of teacher and student model through alignment and uncertainty guided depth selection measurement. To verify the effectiveness of BiCross, we conduct extensive experiments on three scenarios, including Synthetic to Real, Extreme Weather, and Scene Changing. The code and datasets will be released.