How can I diagnose object labeling and OCR problems?

There are three common causes of labeling problems:

  1. User has insufficient budget (vision_services_rate and vision_services_capacity);
  2. The number of salient objects moving in each 12s period > vs_max_per_event;
  3. The stream's video encoding settings need to be adjusted for better motion information.

The first step in diagnosing the problem is to use the Croppings menu item in the Event Viewer overflow menu:

croppings-menu-item-Screen_Shot_2018-03-16_at_10.48.24_AM.png

This displays all the cropped images that were used in visual analysis of the moving objects in the Event in the lower-left corner of the Event Viewer:

cropping-example-back-door-Screen_Shot_2018-03-16_at_11.24.48_AM.png

When you hover the mouse over the cropping thumbnail, it expands to full size:

full_size_cropping_Screen_Shot_2018-03-16_at_11.28.26_AM.png

The number of images per Event is controlled by vs_max_per_event and each cropping uses the full resolution of the video stream unless you've limited the resolution by specifying the optional extraction_height.

If the moving object is poorly cropped, then try optimizing the H.264 encoding by using Variable Bitrate (VBR) while ensuring that the interval between I-frames is > 2 seconds.

Have more questions? Submit a request

Comments