search

virajmavani / semi-auto-image-annotation-tool

Anno-Mage: A Semi Automatic Image Annotation Tool which helps you in annotating images by suggesting you annotations for 80 object classes using a pre-trained model
https://www.virajmavani.com/saiat
Apache License 2.0
577 stars 127 forks source link

Plotting annotated co-ordinates back to orignal images #19

Open Varat7v2 opened 4 years ago

Varat7v2 commented 4 years ago

HI @virajmavani, I have annotated using this semi-automated tool, but I am getting confused with the (xmin, ymin), (xmax, ymax) that I have got after annotating as they won't match on original images. How can I convert it those annotated co-ordinates back to original image dimensions?

virajmavani commented 4 years ago

@Varat7v2 The app converts all the images to match the Pascal VOC format of {width}x500 or 500x{height}. We can add an option to keep original resolution for annotations and re-interpolate the annotations. Re-interpolating might not exactly be the actual annotation that you did. The problem here is essentially the fact that images might be very hard to annotate if they are larger than your screen resolution as they would not fit on your screen. Anyway, images are almost always down-sampled before training and you might want to consider that.

Varat7v2 commented 4 years ago

I tried converting annotated dimensions back to original images by multiplying / dividing aspect ratio of original image with normalized width/height (divided by 500) of a down-scaled images (that we have used for annotating). Though in most of the images bounding boxes were corrected, there were number of images where still bounding boxes were mismatched. I encountered this issue after annotating 4000-5000 images, now I am in trouble as more than 1000 images, I could not solve this issue.