Model Evaluations:

• sky_sdf_1: current experiment that inpaints sky with fixed max-value
• final_2: model without "sky knowledge", i.e. where sky collapses into foreground
• inverse_1: experiment where we used a inverse-depth ("variant") supervision instead of simple log-depth

Fog density in SeeingThroughFog vs. Muses

Using the sparse LiDAR scans, and combining them with some heuristics from DCP and contrast mapping, even the "dense" split showed a much more "clear" split than the fog-split of muses.

The past weeks were about understanding where and why our (metric) diffusion model lacks behind SOTA (but trained on clear weather data) MMDE models, namely UniDepthV2 and DepthAnything3.

The most unintuitive finding was that both SOTA models actually improved in (especially near-field) performance given a dataset with (synthetic) fog vs it's clear counterpart, i.e.:

This week was mostly about evaluating the current model, comparing it naiively with Marigold - always scale&shifting the relative depth estimation output of both models to be comparable with the metric ground truth.

Last two weeks were spent with lots of experiments on the RDS dataset and MERGE architecture.Status quo was:- Dehazing via SD2.1 works well- Novelty from MERGE: Use Marigold adjacent architecture with dual-converters ("prediction heads") for image generation and depth maps.- Early MERGE experiments worked well on VKITTI2 (352 x 1216 RGB)- Extend tests to real-drive-sim