Effectiveness of Batch-wise Weighted Loss for On-GEBD Accuracy

Determine whether the batch-wise loss weighting strategy—multiplying the boundary-target loss by the batch-specific boundary/non-boundary ratio during the training of the Consistent Event Anticipator in ESTimator—yields a noticeable improvement in Online Generic Event Boundary Detection accuracy on the Kinetics-GEBD dataset.

Background

Online Generic Event Boundary Detection (On-GEBD) requires detecting taxonomy-free event boundaries in streaming video without future frames. ESTimator addresses this using a Consistent Event Anticipator (CEA) trained with EST and REST losses. Because boundary frames are much rarer than non-boundary frames, the authors incorporate a batch-wise weighted loss that scales the boundary-target loss by the batch’s boundary/non-boundary ratio to mitigate class imbalance during training.

In supplementary ablations on Kinetics-GEBD, the authors report a 0.5%p Avg. F1 improvement when using batch-wise weighted loss and explicitly conjecture that the observed gain reflects a real accuracy improvement. This raises an unresolved question about whether the observed improvement is genuinely meaningful rather than a trivial fluctuation.