Test-Time Prompt Tuning for Zero-Shot Depth Completion

Least squares fitting (c) Our fine-tuning (d) Our visual prompting (a) Monocular depth estimation

Input RGB Initial Depth Metric-scale sparse depth

Learnable Parameters: 0.9M Learnable Parameters: 363M Figure 1. Comparison of methods for correcting the inaccurate scale of initial depth estimates produced by the depth foundation model using sensor-derived sparse depth with an accurate scale, represented as yellow dots. (a) The depth foundation model generates an initial depth map from a single RGB input. (b) The least squares fitting aligns the initial depth map with the sparse depth [54]. (c) The parameters of the foundation model are fine-tuned under supervision from the sparse depth. (d) Alternatively, the model's parameters are frozen, and a visual prompt is learned to adjust the depth estimates. Please refer to Fig. 3 for depth profiling results and error analysis.