Training & Feedback

Cogniac's Platform can identify any object that a person can reliably visually identify. Sometimes different levels of expertise are required to identify objects or patterns, and therefore generating training data may be more complicated for some applications.

What Training Images Are Most Valuable?

The most beneficial training images are the actual images from your own application domain. Images or videos from other domains or contexts will tend to have different feature characteristics which will make them less useful for training purposes than the actual images you intend to process. For example, if you are processing security camera video, use your actual security camera video for training. Use of internet videos will not necessarily produce as good of results as if you used your own videos for training.

Other guidelines for training images would be to include multiple views of the subject under different lighting conditions and at different angles that will be representative of the expected real-world angles and views.

Feedback Guidelines

It is important to be able to visually identify the object of interest without using other contextual information within the imagery or metadata. The two photos on the left would be excellent training images for 'gun' subject_tags, but the photo on the right would not be a good training image. In this case, the two left images are clear that this officer has tactical equipment including a weapon in a holster on his right side. Unfortunately in the far right image it is impossible to visually confirm that the gun is in the photo - we would need other context or knowledge to 'infer'. The best training images do NOT require contextual knowledge - only visual identification.