Data augmentation for Coco data type- Correlate the Bbox coordinates

UKarthik · February 4, 2020, 4:54am

I am working on my own set of data, I have created coco annotation file for that data I tried train with with unet_lerner with BS=1,I got CUDA memory error, It might be because of of my image size 2448*1564. If I want to resize the image to (256 x any) with Bbox is there any built-in method in fastai to correlate the Bbox coordinates.

muellerzr · February 4, 2020, 4:59am

There are BBox transforms. What CUDA memory error are you getting? Out of memory? Or something like CUDA Assist?

muellerzr · February 4, 2020, 5:03am

Also unet out of the box is for segmentation, not boxes. You’d want something like RetinaNet for that

UKarthik · February 4, 2020, 5:04am

RuntimeError: CUDA out of memory Error, I am with 2080Ti for training

muellerzr · February 4, 2020, 5:05am

Definitely image size. Just scale it down to 224. There are also examples on the forums showing how to do bounding boxes in the framework, do a quick search for RetinaNet for an example with the API

UKarthik · February 4, 2020, 5:09am

I have to scale it to 224 and manually annotate all my data set Right??

muellerzr · February 4, 2020, 5:25am

No you wouldn’t, this is why we have the API

UKarthik · February 4, 2020, 5:28am

Thank you @muellerzr I will check on it