Assigned
Status Update
No update yet.
Comments
gs...@google.com
An application program can simply scan the list of labels, as returned from the call to the Vision API, and so detect the values of interest. Values returned from web detection can be treated similarly. There seems to be little value of implementing this functionality within the API itself. What is the use case(s) supporting the need for this internalized label filtering? A detailed and well-structured use-case, preferably in step-by-step order, usually helps a lot in getting the feature request evaluated quickly, and eventually implemented.
You may refer to the "Detecting Labels" documentation page [1]. The returned element labels = response.label_annotations is immediately useful, if you want to scan a list for relevant elements.
[1]https://cloud.google.com/vision/docs/detecting-labels
You may refer to the "Detecting Labels" documentation page [1]. The returned element labels = response.label_annotations is immediately useful, if you want to scan a list for relevant elements.
[1]
ma...@gmail.com
Naively filtering labels is a terrible approach. If in a picture you want to identify a target, the target quality\visibility could be worse than a non-targeted element and many of its related labels may have a too low confidence level to be returned (and max number of labels you can get is also very limited despite the max results you can set).
I suppose that when neural network have been trained you have trained them by classes and subclasses of images e.g.
animal->mammal->dog->bulldog and something similar hierarchical approach.
Using a keyword you can perform class specific detection that would return much more accurate results since the detection search would be performed among specific class.
I suppose that when neural network have been trained you have trained them by classes and subclasses of images e.g.
animal->mammal->dog->bulldog and something similar hierarchical approach.
Using a keyword you can perform class specific detection that would return much more accurate results since the detection search would be performed among specific class.
gs...@google.com
When you use online prediction, as in the case of the Vision API, the ML service runs a model and returns the requested predictions as the response message for the call. It is unpredictable which labels would get returned as the result of the Vision API call, as each submitted image is usually different from the other. You seem to assume that parameter-dependent prediction is possible and implemented by the Vision API. Do you happen to have more detailed information and practical examples? Is there such a service, that you are now taking as an example?
ma...@gmail.com
each submitted image is different(obviously), but nothing prevent you to use a parametrized class-based detection, you can achieve this with a proper strategy that uses a semantic tree to determine what model checkpoints the system have to use in the request. For this purpose if the model you use has a weak parametrization, It should be trained on all classes and the specific subclasses separately following the hierarchical categorization.
So if you want to detect dog bread, passing that argument the backend doesn't try to perform detection using the global checkpoint of the model trained on all classes of elements but the checkpoint of the specified class of interest. This would dramatical improve the value of Vision service and the accuracy in scenarios that require specific targetization and not only a very generic identification.
So if you want to detect dog bread, passing that argument the backend doesn't try to perform detection using the global checkpoint of the model trained on all classes of elements but the checkpoint of the specified class of interest. This would dramatical improve the value of Vision service and the accuracy in scenarios that require specific targetization and not only a very generic identification.
gs...@google.com
You feature request enjoys now Developers' attention, and will get evaluated soon. There is no decision to implement it, as yet, and no related time-frame. Related developments will get reflected in this thread.
Description
A picture could contain many different elements that are returned and in many use cases developers need to get only info related to some specific topic according the use.
E.g. If I need to create targeted tasks in a picture and only need labels related to clothing to give specific suggestion according the style, and don't need that App list labels that have nothing to do with my purpose, would be great use something like
detection.setTopic("clothing");
in the request to avoid to get results that aren't useful to my needs.
in the same way if someone want to identify info more specific about dogs could set
detection.setTopic("dog");
to retrieve only info related to the dogs and not cars, tree, balls, and whatever thing the picture may contain in addition to the target.