ldkong1205
commented
8 months ago Hi @DachunKai, thanks for asking this interesting question!
To ensure fairness in comparison, we did not include augmentation techniques like de-foggy in our benchmarks. However, we agree that using some de-noising approaches is promising in improving the OoD robustness. Several of our participants in the RoboDepth Challenge have already proven the effectiveness of this two-stage method. You can find more details from the competition report (https://arxiv.org/abs/2307.15061).
Excellent work!
Would it be appropriate to consider comparing it with two-stage methods that incorporate enhancement algorithms for various common corruption types along with depth estimation? For example, when dealing with foggy conditions, could we explore a comparison with the approach of 'defoggy first, followed by depth estimation'?