ETH Computational Intelligence Lab 2025: Uncertainty-Aware Ensemble for Monocular Depth Estimation
Type: Course Project | Year: 2025 | Topics: Computer Vision, Uncertainty Estimation, Ensemble Modeling
In this project, we explored whether monocular depth estimation can be improved by combining multiple predictors in a way that explicitly accounts for model uncertainty.
To test this idea, we fine-tuned a Dense Prediction Transformer variant that predicts both depth and epistemic uncertainty using a Gaussian negative log-likelihood objective. We then trained several experts for different room types and combined them with a separate meta-model that uses the experts’ uncertainty estimates when forming the final prediction.
The resulting ensemble improved siRMSE over the base model and often combined predictions in a way that matched the intended room-specific specialization of the experts. At the same time, the project highlighted practical challenges such as sensitivity to hyperparameters and the risk of mode collapse.