aleju / imgaug

Image augmentation for machine learning experiments.
http://imgaug.readthedocs.io
MIT License
14.43k stars 2.45k forks source link

Support for RGBD Images #429

Open ApoorvaSuresh opened 5 years ago

ApoorvaSuresh commented 5 years ago

I have 4 channel rgbd images of type float32. How does imgaug work on these images? Is the depth channel automatically considered as heatmap and augmented and later concatenated?

Thank you

aleju commented 5 years ago

Most augmenters will treat the fourth channel exactly like the other three channels. Some augmenters will also assume that the fourth channel is A in RGBA and ignore it (i.e. only augment the other three channels). I think only the colorspace-related augmenters do that. Though they will fail for float32 inputs anyways.

The depth map is not automatically converted to a heatmap. Note also in case that you manually convert the depth maps to heatmaps, that the heatmap augmentation is for ground truth data and will hence only apply augmentations that modify the spatial locations of pixels, which might not be what you want.