In the provided dataset, we're given with a '.csv' file for each datasets of UAV frames. It seems to contain information about every frame present in the UAV dataset. How do we generate that information (or the '.csv' file ) for our own dataset? Also, shouldn't we provide the information only for one frame ( initial frame ) of UAV flight? and the model should itself predict information about other frames?

How can I find the ground truth position of UAV frames to calculate the Euclidean distance error? I tried this on the village dataset.

Hi, I have a question about how you're computing p_{M,k} and also the corner loss during training. Specifically, if we're using a CNN with say a 2x2 max-pooling layer, then the output feature maps (i.e. channels) will effectively be resized to half the dimensions of the original image. And so I would expect we'd have to rescale the ground-truth warp parameters accordingly before computing the corner loss during training, but it doesn't seem like that's actually happening anywhere. Similarly, there also doesn't seem to be any rescaling going on when computing p_{M,k}.

Am I misunderstanding what's going on? I think I might be confused as to the relationship between warp parameters in image space and those in feature space. Would very much appreciate if you could help clarify.

Running the code with the recent versions of the libraries gives me an error

AttributeError: 'function' object has no attribute 'Variable'

in DeepLKBatch on the lines like

if isinstance(p, torch.autograd.variable.Variable):

The problem is fixed by converting the reference to torch.autograd.Variable. See lanpa/tensorboardX#100

Even if the original authors are not going to fix it, I hope the information will help those who try to use their code.

Hi! I've noticed that some codes might be missing, which are 'decompose_rel_hmg.py' and 'combine_relative_pose_small_seq.m' mentioned in the comments. I have tried to convert the motion parameters to GPS coordinates by simply decomposing the homography matrixes between the village map and the UAV images in origin resolution (4608*3456). However, the results are not good as they should be.

What goes wrong with it? Would you please show how the homographies are converted? Look forward to your reply.

I want to run this code with different data. Hence, i should find initial motion of image sequences (frames). I found 4 satellite map points in GIMP which corresponds to UAV frame. I replaced homography values in warp_hmg(img, p) with cv2.getPerspectiveTransform() homography as an initial motion. cv2.warpPerspective() works well, but when i put my H to "xy_warp = H.bmm(xy)", it gives a black image. How can i calculate initial motion.

Note: i tried this experiment in Village.

Could you kindly check the status of the link you provided?
Thanks in advance,