r00tman / eventnerf Goto Github PK

could you provide "genevents_color_altbg.py"? I want to create my own event data to test

How to use other event camera datasets (like https://daniilidis-group.github.io/mvsec/ and https://dsec.ifi.uzh.ch/) as model input?

Nice work, I have a question about the training.
In your paper, you train the model for 5 * 10^5 iterations. However, N_iters in config.txt is set to 5000001. What is the correct number of parameters?

Hi, I have two questions regarding the real sequences.

• Alignment with events and RGB images.
  When I load data/real/legocube/train for example, events and RGB are not aligned.
  Assume the RGB file names correspond to the millisec unit, (ex. r_00000.png = 0sec, r_00001.png = 0.001sec),
  An example output of the same timestamps from event camera and image is as below and looks not properly aligned:

(Please ignore the wrong color of the RGB, it's basically because of my visualizer. What I want to ask is the alignment between frame data and event data.)

• The image files in other sequences:
  The RGB files from other real sequences (I checked at least plant, chick, mic) look like just legocube. Can you double check if you upload image data correctly?
  

Or am I doing anything wrongly?

Best,

Shintaro

This is a awesome work! But I meet some trouble that I can't resolve.

Collecting package metadata (repodata.json): done
Solving environment: failed

ResolvePackageNotFound:

• ipython==7.30.0=py39hf3d152e_0

Hi
I tried scripts: job_nerf_lego.sh and job_nerf_drums.sh with V100 GPU, after 6 hours the psnr was still around 10 I wondered did I do anything wrong. How can I get the number from the table on paper?

Best

Hello, thank you very much for your great work.
After training for 500,000 times, I tested using the command you provided, which is as follows:
python ddp_test_nerf.py --config configs/nerf/lego.txt --render_split train --testskip 10

But the result is not good

Hello, I reproduced your experiment, currently testing on the lego dataset, all the parameters have not been touched, but I found that even if I iterate to 1 million times, I can't achieve the effect of PSNR25 in the paper, in fact it is only about 17, why is that? The parameters are kept the same as in your github repository.

Hi, thanks for the great work.
I have noticed that the real-world data file that has been released does not include the paired RGB data, which was utilized in the paper to assess deblur-nerf. Is it feasible to make this portion of the data available? It would be immensely beneficial for evaluating other nerf-based techniques on this dataset.

Hi, thank you for the dataset.

When I load the event npz file, the timestamp looks strange (all integers). What's the actual unit of the timestamp?

import numpy as np
fname = '/home/ubuntu/slocal/ssd_data/EventNeRF/data/lego/test1_events/test/events/out.npz'

with np.load(fname) as data:
    ts = data['t']

ts[:10]     # returns array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0])
ts[-10:]    # returns array([999, 999, 999, 999, 999, 999, 999, 999, 999, 999])
len(np.unique(ts))  # returns 1000
ts.max()  # returns 999

I suspect that this integers [0-999] corresponds to the number of frames.
If that is the case, this dataset does not contain accurate event timestamps, but rather has the ones quantized to the frame IDs?

in my docker container, DDP training is stopped when the create_nerf() is called. Do you know what is the problem about it?
I used 8 gpus and I tried extract mesh and it is successfully done. (but trainning is stopped!!)


Do you know why DDP is stopped? (is there any deadlock..?)