DiffPose icon indicating copy to clipboard operation
DiffPose copied to clipboard

custom dataset load for train

Open ryukyungjoon opened this issue 1 year ago • 1 comments

HI, Vivek

There has been a lot of progress in developing with your help. Thank you.

I have some question about custom dataset loading for train. My Dataset's DRR Image size is 160x52(width x height) And proj-params/pixel-col-spacing, proj-params/pixel-row-spacing parameter is 1.0 not 0.194. The results of the model trained with my data do not seem to be good. It is the process of finding out how to improve it.

image

Model Evaluation result is this: 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49 specimen_01_best,1384.3680419921875,1401.7911376953125,1407.7540283203125,1398.0103759765625,1406.6297607421875,1403.4468994140625,1402.82861328125,1398.331298828125,1407.64892578125,1406.0977783203125,1401.754638671875,1404.6663818359375,1408.11669921875,1404.620849609375,1407.6368408203125,1405.5894775390625,1409.061279296875,1401.509033203125,1409.2413330078125,1411.951171875,1412.0123291015625,1406.8592529296875,1410.77099609375,1417.6480712890625,1410.14013671875,1414.9044189453125,1409.55615234375,1413.5372314453125,1403.1396484375,1406.974365234375,1406.1654052734375,1401.591552734375,1409.28466796875,1408.3173828125,1412.5299072265625,1397.3001708984375,1398.3497314453125,1399.67138671875,1394.5645751953125,1406.7659912109375,1410.5157470703125,1412.334228515625,1392.8814697265625,1395.0009765625,1393.5361328125,1393.4935302734375,1398.77734375,1400.831298828125,1409.1187744140625

I need your help Vivek. Thank you.

ryukyungjoon avatar Jun 24 '24 03:06 ryukyungjoon

what is mean the subsample?

ryukyungjoon avatar Jun 27 '24 01:06 ryukyungjoon

subsample is how much you downsample the image. E.g., the original images are 1436 x 1436 px (the original height is 1536 but we subtract 50 pixels from each edge to remove the black border). It's computationally slow to render DRRs that large for iterative optimization. Instead, can downsample the image from 1436 -> 256 px. That ratio, 1436 / 256, is the subsample factor.

eigenvivek avatar Sep 20 '24 13:09 eigenvivek