Using the repo on custom data set

[ann-mary-fna]

Hi,

Can you explain how to use this repo on a custom dataset? ( the commands, input data structure, etc ) I wasn't able to understand it fully from the readme.

Thanks.

[huangjian39]

Hi,

Thanks for your question! Here’s how you can use this repo with your own dataset:

Step 1:
Organize your custom dataset in the same structure as our Replica (event) dataset.
You can find detailed instructions in Section 5. “Dataset Format” of the README.

Step 2:
Run the following command:

python main.py --config configs/SimuEvent/replica_room0.yaml
(You may need to adjust the file paths or the YAML file name according to your setup.)

Step 3:
Fine-tune the parameters in the YAML configuration file to fit your dataset.

Hope this helps!

[ann-mary-fna]

Don't we also need to change "datadir" ( to the path to the input folder ) as well as "output" variable in the yaml file used as config ?

[huangjian39]

Yes, you need to set those parameters according to your own dataset.
What I mentioned was just the general procedure — you’ll need to handle the specific code setup yourself.

[ann-mary-fna]

Thanks for clearing the doubts.

I need to use this repo on images taken from a normal camera. We were thinking of putting the images in the event_threshold_0.1 folder. Will this be the correct approach? If so, should we simulate events from these images in dataloader.py file ?

[huangjian39]

Sorry for the late reply — I’ve been quite busy over the past few weeks.

To clarify: the event_threshold_0.1 directory is meant to store event streams, not RGB images. So placing raw RGB frames directly in that folder would not work.

From your description, it sounds like you are trying to use normal RGB photos and simulate event data. If that’s the case, here are a few suggestions:

• Upsample your RGB sequence

A standard camera might capture at ~20–30 Hz. For event simulation, you usually need much higher temporal resolution (e.g., 500 Hz or even 1000 Hz). You can interpolate intermediate frames (e.g., using frame interpolation networks).

• Use an event simulator

After interpolation, you can convert the high-frame-rate RGB sequence into an event stream using an event simulator.
I previously used this one from the RPG group:
https://github.com/uzh-rpg/rpg_vid2e

• Prepare the dataset structure

Once your event stream is generated, you can place it into the corresponding event folders following the dataset format described in the README (Replica-style layout).

This pipeline is a bit involved, and you may need to carefully handle interpolation quality, timestamp alignment, etc., to get reasonable event simulations.