search

alexzzhu / EventGAN

Official repository for the paper: Zhu, Alex Zihao, et al. "EventGAN: Leveraging Large Scale Image Datasets for Event Cameras." arXiv preprint arXiv:1912.01584 (2019).
MIT License
67 stars 10 forks source link

Question about the generator output. #2

Closed YongjianDeng closed 4 years ago

YongjianDeng commented 4 years ago

Hi,

I'm wondering if your generated fake event in the form of volume but not point cloud? It is like you are trying to achieve the volume-like event representation but not the raw events[x, y, t, p] form, right?

Best, Ty

alexzzhu commented 4 years ago

Right, the network is predicted an event volume, rather than discrete events as a point cloud.

YongjianDeng commented 4 years ago

Hi,

Sorry to spare you so much time. After the experiment for optical flow prediction. I found that the ground truth is far from the prediction results, and the prediction results outperform the ground truth in terms of the images' warping task. Did you encounter this situation or is there something wrong with my setup? Does that mean the ground truth with lower accuracy compared to the prediction results trained using warping loss? If it is, how can we measure the optical flow quality using this dataset?

Ps: Before comparing, I have already transform the predicted flow from the range [-1, 1] to pixel level flow just like the ground truth is.

Best, Ty

alexzzhu commented 4 years ago

No problem! How exactly are you measuring the error wrt to the ground truth? And how different are the losses? Please note that the ground truth timestamps do not align with the image timestamps, and so some interpolation is necessary. Please see this script for an example of how we compute the errors: https://github.com/daniilidis-group/EV-FlowNet/blob/cf801ef5e4b112e08bfe6eac4f0878b1f5aac2fe/src/test.py#L107

YongjianDeng commented 4 years ago

Hi,

Thx for your timely response. I did follow this script to measure the error. However, from the visualization result, warping results from the predicted sparse flow better than the warping performance using the gt flow. My question is if in some situations, predicted flow is more accurate than gt flow?

Best,

alexzzhu commented 4 years ago

That's definitely possible. The ground truth was measured with egomotion+depth, and so there may be errors due to error in either measurement.

YongjianDeng commented 4 years ago

I appreciate your patient response, which helps me a lot.