search

chan8972 / Spike-FlowNet

MIT License
98 stars 17 forks source link

Pretrain Images #3

Closed 00Sim closed 3 years ago

00Sim commented 3 years ago

Hello, thank you for sharing your code and great work.

I am trying to view the results of the pretrained checkpoint_dt1.pth.tar and checkpoint_dt4.pth.tar model on the indoor_flying1

But it seems that the Spike Image and Predicted Flow Output visualizations are very sparse, even with the 4 frame model (results for checkpoint_dt4.pth.tar on indoor_flying1 shown below). PreTrained-4Frames

I am running with the existing parameters in the code: batch_size_v = 4, sp_threshold = 0.5

Should any of the parameters be changed to get a denser spike input and thus denser predicted flow output?

chan8972 commented 3 years ago

-Typically, the spike events are very sparse at the beginning of the video and become denser when the drone starts to fly and roam around the room environments in indoor_flying data. -Depending on your computer specs, the code may run slow (taking at most an hour to finish) to test the model and visualize the images/flows. -Please set batch_size_v=4 and sp_threshold=0.5 for dt=4 case and batch_size_v=4 and sp_threshold=0.75 for dt=1 case. -If the problem remains for entire video sequences, please check whether pre-trained models are loaded appropriately. Thank you for your interest and let me know if this solves your issue.

00Sim commented 3 years ago

Ah got it, the visualizations look as expected later on in the video as the spikes become more dense. Thank you!