Browse Prior Art Database

Method and apparatus for cross-view correspondence identification Disclosure Number: IPCOM000241616D
Publication Date: 2015-May-18
Document File: 2 page(s) / 221K

Publishing Venue

The Prior Art Database


A method for matching identical person or objects from multiple visual images from cameras in different places, There are two main components: 1) Learning based patch correspondence structure estimation for camera pair, which further consists of a) Building patch set with structure; b) Learning to set the value for the patch correspondence map 2) One-to-one assignment based patch matching and overall scoring, which further consists of a) Combine both learned correspondence structure and patch visual similarity to set the patch-to-patch similarity for all pairs; b) Impose one-to-one patch matching to obtain the ideal matching and associated overall score

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 100% of the total text.

Page 01 of 2

Method and apparatus for cross

Method and apparatus for cross-

--view correspondence identification

view correspondence identification

The problem this disclosure addresses is to recognize identical people from multiple cameras views. It has many applications such as in video surveillance.

The system overview of our approach is as follows:

The first component is learning the correspondence structure. The embodiment is as follows:

1) Set up the patch set by a certain order

2) Perform patch based matching such as [1] to find a reasonable mapping M (not perfect) between the probe sample and one of the positive samples (usually with different poses)

3) Compute the score S of each matching pair between the probe image and one of the positive samples from gallery images


Page 02 of 2

4) Add up each S*M to obtain the learned correspondence structure given k-1 positive samples

The second component is robust matching via one-to-one constraint

1) Best score matching w/o one-to-one constraint, cannot avoid one-to-many matchings (learned correspondence structure also allows for one-to-many matching to enhance flexibility)

2) One-to-one constraint imposed, more reasonable matching results and more realistic overall matching score

Perform Hungarian method to find the one-to-one patch matchings, and then compute the overall patch matching score as the similarity score between the probe and one gallery image