RuntimeWarning: Mean of empty slice ... or ... gdalbuildvrt ERROR 1: No input dataset specified. about s2p

[absudabsu]

I tried with a large batch of images, which are guaranteed to overlap. No luck, same error.

Looking at it again... Perhaps rather np.mean() returning NaN (a warning), the real error is: ERROR 1: No input dataset specified. ? I'm not sure how what I'm doing wrong.. Any hints?

[dyoussef]

Hello @sirgogo,
Thank you for bringing up the problem to our attention.
Please, can you send your dataset?
Thanks in advance

[absudabsu]

@dyoussef Thanks for the reply. The dataset I am using is available here: http://www.jhuapl.edu/satellite-benchmark.html . From my understanding, its the same dataset used in the IARPA top coder challenge, but I could be wrong. The specific images I am using are the ones available in the (above, specified) dataset, such as: iarpa_3D-satellite-dataset/Challenge_Data_and_Software/cropimagedata/Explorer/...

I did some more digging, and I think that maybe the issue is due to images not overlapping 100%, since I found that for some image pairs (just two at a time), the algorithm does run, but tries to overallocate memory (I see all kinds of malloc calls on the screen) when it is computing the global DSM. However, for other image pairs, it fails arbitrarily like above.

I tried a similar strategy on a smaller reconstruction are by setting 'full_img' = True and specifying a small ROI. For some image chips the reconstruction was able to complete (albeit on a very small area)! For other chips, it fails arbitrarily like above.

I have not inspected thoroughly, but from my initial experiments, it seems like the algorithm works/runs when two supplied images are very close (e.g. consecutive), such that their projections cover the same/similar geodesic bounding box with very high overlap, and their image-plane coordinates are also simliar.

[carlodef]

Hi sirgogo,
Could you please specify the filenames of the input images and provide your full json configuration file so that we can try to reproduce the failure?

From your output I see that the size of your input images is about 2000 x 2000 pixels (from 286 * 7). The original images of the IARPA challenge are much bigger. Maybe your issue comes from the fact that you've cropped the input images without modifying the corresponding RPC coefficients. There's a tool for that: rpc_cropper.