Daniel Koguciuk

@irfan798 - works also for TF1.14 - thanks! :heart:

I was just going to describe the exact same problem. I've downloaded raw data given the link here: https://github.com/scaleapi/pandaset-devkit/issues/67 I've read the raw point cloud in lidar coordinate frame like...

@xpchuan-95 , sure, I am ok with motion compensation :) I would be even more ok if the rigid transformation from ego frame to lidar frame would be provided by...

@ltriess could you have a look at it? It's related to #76

Ah, and a use case: Previously, this code gives first iteration over the whole dataset, then no iteration, since the information about the sequences was gone: ``` for sequence in...

@rmcavoy, could you please elaborate? What do you understand by levels 3, 5 and 7? I've read the original implementation and they have both **1x1 conv2d and maxpool** on **both...

I see your point. In the EfficientNet paper, they call it stages (Table 1), so I believe you're referring to the 4th, 6th and 8th stages which are then treated...

Yes, totally agree. And again yes: I've double-checked it and you're right: there is no 1x1 Conv2D between P6 and P7 - my bad!

> Also, Figure 3 shows each layer of the box/classification network having two (depth-wise separable) convolutions per layer as they surround the two convolutions with a dotted box just as...

@rmcavoy, don't worry - you've help a lot of people here with your comments :)