FFS underperforms vs FoundationStereo on calibrated data #44
Description
Hi @wenbowen123, thanks for making this work publicly available.
We are testing both models on a custom-built, well-calibrated stereo setup (not an off-the-shelf system). FoundationStereo performs very well in terms of both scene geometry and metric accuracy. However, on the exact same data, Fast-FoundationStereo produces unusable results.
Do you have any recommendations on how to diagnose this discrepancy? One possibility we are considering is a domain (covariate) shift between the data used for training/distillation and our setup, which may be affecting generalization. Another possibility is that the compression/distillation in Fast-FoundationStereo reduces robustness in certain real-world calibrated scenarios.
If there are known sensitivities, recommended preprocessing steps, or inference settings that could improve performance, we would really appreciate your guidance.