StridedSequenceDataset target indexing

[maycherryblossom]

Hi, I’ve been working with your RoNIN model (ResNet) and had a question regarding data indexing for feature and target during training.

In the paper, it mentions fitting the velocity at the i-th frame with features from frames i-200 to i. However, in the provided code, particularly in StridedSequenceDataset, the dataset indexing seems different:

feat = self.features[seq_id][frame_id:frame_id + self.window_size] targ = self.targets[seq_id][frame_id]

This appears to fit the velocity at the i-th frame with the features from frames i to i+200, rather than i-200 to i as described in the paper. Could you clarify how the data is being indexed and why there seems to be this difference?

Thanks!