Non-maximum suppression of features for object detection

A non-transitory computer-readable storage medium having stored thereon data representing sequences of instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:

• identifying and evaluating by a processing system a plurality of features in a received image for object detection, the image including one or more objects, the processing system including a graphical processing unit (GPU);
• dividing the image into a grid of cells and generating a learnable non-maximum suppression threshold for each cell of the grid;
• processing each detected feature in the image for non-maximum suppression, wherein processing of each detected feature includes:
• determining which cells the feature is located within and assigning the feature to the determined cells, and
• comparing the feature in the determined cells with any previously selected features to remove duplicate cells, including performing parallel processing for each feature according to each of one or more cells the feature is located within, the processing including applying the non-maximum suppression threshold generated for each of the one or more cells;
• and
• detecting an object within the received image based at least in part on the parallel processing of each feature identified in the image.

A processing system comprising:

• a central processing unit (CPU);
• a graphical processing unit (GPU), the GPU being capable of a plurality of processing threads;
• dynamic random access memory (DRAM) to store data for object detection, including data for detected features within a received image;
• and
• an object detection algorithm including a non-maximum suppression element;
• wherein the object detection algorithm of the processing system is to provide for non-maximum suppression of features that are detected in an image, including the processing system to divide the image into a grid of cells and generate a learnable non-maximum suppression threshold for each cell of the grid, and to provide parallel processing for each feature according to each of one or more cells the feature is located within, including:
• determining which cells the feature is located within and assigning the feature to the determined cells, and
• comparing the feature in the determined cells with any previously selected features to remove duplicate cells, including performing parallel processing for each feature according to each of one or more cells the feature is located within, the processing including applying the non-maximum suppression threshold generated for each of the one or more cells;
• and
• wherein that processing system is to detect an object within the received image based at least in part on the parallel processing of each feature identified in the image.