PROJECT TITLE :

Implicit Negative Sub-Categorization and Sink Diversion for Object Detection - 2018

ABSTRACT:

In this Project, we tend to specialize in improving the proposal classification stage in the object detection task and present implicit negative sub-categorization and sink diversion to elevate the performance by strengthening loss perform in this stage. 1st, based mostly on the observation that the “background” category is usually very numerous and therefore challenging to be handled as one indiscriminative class in existing state-of-the-art methods, we propose to divide the background class into multiple implicit sub-categories to explicitly differentiate various patterns among it. Second, since the bottom truth category inevitably has low-value probability scores for certain images, we have a tendency to propose to add a “sink” category and divert the possibilities of wrong classes to this category when necessary, such that the bottom truth label can still have a better likelihood than different wrong classes although it's low probability output. Additionally, we have a tendency to propose to use dilated convolution, which is widely used in the Semantic Segmentation task, for economical and valuable context info extraction. Intensive experiments on PASCAL VOC 2007 and 2012 knowledge sets show that our proposed strategies based mostly on faster R-CNN implementation will achieve state-of-the-art mAPs, i.e., 84.onepercent, eighty two.half dozen%, respectively, and get 2.5p.c improvement on ILSVRC DET compared with that of ResNet.