Closed mitkina closed 5 years ago
Hmm, is this the default model on KITTI? And doing t+1 prediction for each time step? Have you tried training it up to 20 timesteps or is this just in inference mode?
This is not quite on the default KITTI data, but comparable. Doing t+1 works just fine - it's the actual longer term prediction (e.g., recursively predicting for 2 seconds from an input of 0.5 seconds) that deteriorates pretty quickly. I perform the fine-tune as suggested.
Thanks!
Ah okay yeah it's probably just that the data gets non-deterministic as time goes on. Would probably need some type of GAN loss for longer term prediction.
Got it. Do you have any context for how far you could push the prediction with the original KITTI data?
I would say up to ~5 time steps it can still be okay. There are a few examples at the end of the page here: https://coxlab.github.io/prednet/. Figure 7 in the paper also has some quantitative info.
Thanks! I was wondering if you tried pushing it any further than 0.5 second predictions. But I guess that's just the limit of the model! Thanks again for your help!
Ah okay yeah it's probably just that the data gets non-deterministic as time goes on. Would probably need some type of GAN loss for longer term prediction.
What you are talking about is using the discriminator network as a loss function or using the GAN network to get some loss function. Thanks for your reply.
Thanks! I was wondering if you tried pushing it any further than 0.5 second predictions. But I guess that's just the limit of the model! Thanks again for your help!
No problem! Yeah there isn't a time step where it suddenly jumps up in MSE (it increases pretty linearly like Fig 7 suggests), but ~5 timesteps is a good rule of thumb.
Ah okay yeah it's probably just that the data gets non-deterministic as time goes on. Would probably need some type of GAN loss for longer term prediction.
What you are talking about is using the discriminator network as a loss function or using the GAN network to get some loss function. Thanks for your reply.
Yeah it would be more along the lines of using a discriminator to add to the loss function.
Do you have any advice as to how to get the network to make longer term predictions (e.g., 2 seconds at 10 Hz) without completely deteriorating by the end?
Thanks!