search

eigenvivek / DiffPose

[CVPR 2024] Intraoperative 2D/3D registration via differentiable X-ray rendering
http://vivekg.dev/DiffPose/
MIT License
118 stars 14 forks source link

custom dataset load for train #29

Closed ryukyungjoon closed 2 months ago

ryukyungjoon commented 5 months ago

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 commented 4 months ago

what is mean the subsample?

eigenvivek commented 2 months ago

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.