Critical bug: View sampling incorrectly limits max_gap to min_gap

[Rvosuke]

Thanks a lot for the wonderful code. But there is a critical bug in the view sampling logic that severely impacts the model's training effectiveness and generalization ability.

Location

• File: src/dataset/view_sampler/view_sampler_bounded.py
• Line: 62

Incorrect Code

max_gap = min(num_views - 1, min_gap)

Correct Code (as per the base library pixelsplat)

max_gap = min(num_views - 1, max_gap)

Issue

The purpose of this line is to cap max_gap to ensure it does not exceed the maximum possible gap between frames in the video (num_views - 1). However, due to a likely typo, min_gap is used instead of max_gap, causing max_gap to be incorrectly set to min_gap. This error restricts the frame gap during training to min_gap, preventing the model from being exposed to larger gaps.

As a result, the model is not trained on a diverse range of frame differences, such as those in the test set (where input frame differences range between 45 and 135). This leads to poor generalization performance on the test set.