0

Screenshot

I have this huge data set for which I have only taken a sample from that data to show you,now as you can see it has two class cat and dog where in the training data i have to label it manually since the cat and dog images are mixed, so is there any alternative way to do it.I have to annotate this then only i can train as to if whether cat or dog.

Dave
  • 5,108
  • 16
  • 30
  • 40
S L SREEJITH
  • 63
  • 8
  • How accurate must the labelling be? If you can have some error rate, then you could always label automatically with an available deep network; many deep networks are quite accurate for dog and cat images, since these are often well represented in common training datasets. – Mozglubov Nov 15 '18 at 21:35
  • Hi Mozglubov, in order to model as from the training set i have to split this as cat and dog first and for larger images in a directory i cannot manually annotate each one as cat.1,cat.2 or dog.1,dog.2,so is this there any other alternative for this or i have to manually do this? – S L SREEJITH Nov 16 '18 at 10:31
  • Well, what I was trying to get at is whether or not you need that exact annotation. Depending on what you plan to do with this dataset, it is sometimes sufficient to have labels which are *mostly* correct. This is sometimes referred to as "soft" or "noisy" labelling. This paper by Reed et al. (https://arxiv.org/abs/1412.6596) provides a good example of achieving strong results despite not having fully accurate training labels. If you absolutely do need a human-accurate set of annotations but cannot do it yourself, then you may want to use services like Amazon's Mechanical Turk. – Mozglubov Nov 16 '18 at 14:29

1 Answers1

0

One possible solution is to upload your dataset to labelbox (link: https://www.labelbox.com/) there you are able to annotate your dataset and then download the results for instance as a JSON file. The web page correlates your images with the labels and then you can use those informations for your work.

Woodstock94
  • 196
  • 10