Differences Between REVIEW and RESEARCH Data
There may be some differences between the number of objects detected in the REVIEW and RESEARCH modules. mainly because the use case for each of these modules is different. REVIEW, used for investigations, may include an object in more than one classification value to ensure the investigator will not miss anything, whereas RESEARCH is geared toward accurate and consistent presentation of the number of people or cars, therefore only counts each object once.
Differences in Applying Filters
REVIEW data is based on probability data, which means that an object can appear under multiple attribute values. For example, if an object is detected wearing a shirt with a red and blue pattern, the object will appear in the REVIEW solution when filtering to both red and blue.
RESEARCH data is based on the maximum value of all object attribute probability values. In the example described above, if the main color of the shirt detected is red, the object will be associated only with the color red in the RESEARCH solution.
In another example, if an object has been classified with 40% probability as a Man and 60% probability as a Woman, the object will be included when objects are filtered to both the Man and Woman classes in the REVIEW solution (as long as it passes the tolerance threshold). In the RESEARCH solution, that same object will only be included when objects are filtered to the Woman class.
Therefore, filter results may differ between the REVIEW and RESEARCH solutions.
Differences in Accounting for Multiple Appearances
In the REVIEW module, every appearance of a person is shown as a different thumbnail to provide the investigator with all relevant information for the investigation. When this data is transferred to the RESEARCH module, another mechanism, called Re-ID (see Tracking People Across Cameras Based on Face Recognition), is applied to link the appearances of a person into a single identity. This is intended to ensure that people counting is accurate and consistent.