Jumpat
Why is the flash ignored when splitting the camera in lerf dataset?
This roots in data distribution. SAM tends to segment the camera into multiple parts, and thus the learned affinity features inherit this trait.
Thank you for your reply. So this shows that even if we introduce scale features, there is no way to solve this problem？ If there any strategy to overcome this issue?
Yeah, rebalancing the positive and negative terms of the loss function can help (but may introduce other problem like segmenting different objects into one). The scale parameter is introduced to alleviate such dilemma. However, whether it can achieve the predefined goal still depends on whether there are enough training sample to adjust the affinity features to record the corresponding segmentation pattern. The best way to fix it is to adjust the hyper parameter in SAM auto mask extractor.