Analyze job allows model performance analysis by comparing the ground truth and predictions and presenting this in a plot where points cluster together based on the model performance characteristics (e.g., points with similar model error form a cluster). A similarity search can be run in two modes for such a job.
- Default: The search looks for the same images that match the model performance characteristics of the search candidate set images.
- Content features: The search looks for the same images that match the visual content of the search candidate set images.
In the default mode, where 'Content Features' is toggled off, the search results show images with similar model performance characteristics. In the below example, the search returned those images with a horse being predicted without it being present in the ground truth annotation.
When 'Content Features' are turned on, it can be seen that the search results are returning visually similar images with the horse jumping the barrier rails, as shown below.
When 'Content Features' is turned on, if the underlying dataset was created with a 'Patch' featurizer selected, you will get an option to do a patch feature based search.