View source for 20240054775. SYSTEM AND METHOD FOR DOMAIN ADAPTIVE OBJECT DETECTION VIA GRADIENT DETACH BASED STACKED COMPLEMENTARY LOSSES simplified abstract (CARNEGIE MELLON UNIVERSITY)

=SYSTEM AND METHOD FOR DOMAIN ADAPTIVE OBJECT DETECTION VIA GRADIENT DETACH BASED STACKED COMPLEMENTARY LOSSES=


==Organization Name==

[[:Category:CARNEGIE MELLON UNIVERSITY|CARNEGIE MELLON UNIVERSITY]]


[[Category:CARNEGIE MELLON UNIVERSITY]]

==Inventor(s)==

[[:Category:Zhiqiang Shen of Pittsburgh PA (US)|Zhiqiang Shen of Pittsburgh PA (US)]][[Category:Zhiqiang Shen of Pittsburgh PA (US)]]

[[:Category:Harsh Maheshwari of Pittsburgh PA (US)|Harsh Maheshwari of Pittsburgh PA (US)]][[Category:Harsh Maheshwari of Pittsburgh PA (US)]]

[[:Category:Marios Savvides of Pittsburgh PA (US)|Marios Savvides of Pittsburgh PA (US)]][[Category:Marios Savvides of Pittsburgh PA (US)]]

==SYSTEM AND METHOD FOR DOMAIN ADAPTIVE OBJECT DETECTION VIA GRADIENT DETACH BASED STACKED COMPLEMENTARY LOSSES - A simplified explanation of the abstract==


This abstract first appeared for US patent application 20240054775 titled 'SYSTEM AND METHOD FOR DOMAIN ADAPTIVE OBJECT DETECTION VIA GRADIENT DETACH BASED STACKED COMPLEMENTARY LOSSES

==Simplified Explanation==

The patent application describes a detach strategy to suppress the flow of gradients from context sub-networks through the detection backbone path, resulting in a more discriminative context by forcing the representation of context sub-network to be dissimilar from the detection network.

* A sub-network generates context information from early layers of the detection backbone.
* Representations from instance and context should be discrepant as they focus on different parts of an image.
* A stacked complementary loss is generated and backpropagated to the detection network.

== Potential Applications ==
* Object detection and recognition systems
* Image classification algorithms
* Video analysis and surveillance technologies

== Problems Solved ==
* Enhances the discriminative power of context information in object detection
* Improves the accuracy of object recognition in complex scenes

== Benefits ==
* More accurate object detection and recognition
* Enhanced performance in challenging visual environments
* Improved efficiency in processing large amounts of visual data


==Original Abstract Submitted==

disclosed herein an effective detach strategy which suppresses the flow of gradients from context sub-networks through the detection backbone path to obtain a more discriminative context by forcing the representation of context sub-network to be dissimilar from the detection network. a sub-network is defined to generate the context information from early layers of the detection backbone. because instance and context focus on perceptually different parts of an image, the representations from either of them should also be discrepant. in addition, a stacked complementary loss is generated to and backpropagated to the detection network.



[[Category:G06V10/82]]
[[Category:G06V10/77]]