Request for access to raw event data — many HDF5 sequences appear heavily subsampled or missing
Hi @GabrieleMagrini ,
First of all, thank you for sharing the FRED dataset — it’s a valuable resource for the neuromorphic vision community!
While preprocessing the event data, I noticed that many sequences in the provided HDF5 files seem to be severely incomplete or significantly subsampled. For instance, in the training split, sequences such as 0, 1, 100, 101, and 102 contain only a small fraction of events compared to what would be expected from a typical neuromorphic sensor recording. This makes it challenging to perform low-level event-based processing or develop methods that rely on the full spatiotemporal structure of raw events.
Would it be possible to share the original, unprocessed event streams (e.g. full .hdf5 format) for these sequences? Access to the raw data would greatly benefit research efforts focused on direct processing of neuromorphic signals—such as event-RGB fusion, spike-based representation learning, or drone detection using asynchronous event cameras—which heavily depend on the integrity and density of the event stream.
I completely understand if there are privacy, storage, or licensing constraints, but any additional data or clarification on the preprocessing pipeline would be immensely appreciated.
Thanks again for your work and support!