Open z157181773 opened 6 months ago
Hi,
In single-thread mode, the tracking process pauses until the mapping completes a certain number of optimization iterations. This slows down the system, but improves accuracy by tracking against a more optimized Gaussian map. This mode is effective in synthetic datasets like Replica,where the input images and depth maps are noise-free, and more iterations with high-resolution images simply lead to better results if don't care about the runtime.
I personally believe this isn't a realistic scenario, as the best strategy here would be running optimization with as many pixels/iterations as possible. Thus, our paper mainly focuses on multi-process results and uses single-thread results just as a reference for benchmarking under similar conditions for comparisons.
Best,
For the dataset replica, each scenario has two configs. I know * _sp. yaml is set to single_thread=True, but why does setting it to single_thread result in so much better results? Why is this