ETH Computational Intelligence Lab 2025: Uncertainty-Aware Ensemble for Monocular Depth Estimation 2

In this work, we explored whether the per- formance of the state-of-the-art models on the Monocular Depth Estimation task can be im- proved by averaging multiple models’ predictions based on the model’s uncertainty. To verify this assumption, we fine-tuned a Dense Prediction Transformer model modified to predict uncer- tainty using Gaussian Negative Log Likelihood loss. Then, we trained multiple experts by further fine-tuning the aforementioned model for differ- ent room types. To combine their predictions, we trained a separate meta-model that takes the mod- els’ uncertainty into account. We showed that this approach decreases siRMSE compared to the performance of the base model. Additionally, we identified that this approach is sensitive to hyper- parameters and can suffer from mode collapse. Finally, our results show that the meta-model reg- ularly combines the ensemble’s individual predic- tions in a manner consistent with the intuition of the expert’s room-type domain knowledge and generalizes to ambiguous images.

The GitHub repository is at Computational Intelligence Lab 2025 Project Repository and the corresponding report at Computational Intelligence Lab 2025 Project Report