This repository contains the reproducible implementation of the three-branch RGB-Event training and dynamic hierarchical model merging pipeline.
-
bash train.shtrains the frame, stream, and voxel branches with classification, logit-level mutual distillation, feature-level mutual distillation, and relation preservation. -
bash merge.shmerges historical full checkpoints with frequency-aware branch experts. Set the checkpoint variables before running it. -
CHECKPOINT=... bash test.shevaluates the merged checkpoint with fused logits.
Before training, set these environment variables for the dataset.