Laplacian Object
See more details and examples in the following papers:
Sujoy Kumar Biswas, Peyman Milanfar, Laplacian Object: One Shot Object Detection with Locality Preserving Projections , IEEE International Conference in Image Processing (ICIP), 2014
Sujoy Kumar Biswas, Peyman Milanfar, One Shot Detection with Laplacian Object and Fast Matrix Cosine Similarity IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015
LSK tensor descriptors are projected onto leading principal components to yield decorrelated and discriminatory feature tensors which are then used in a max-margin training framework with the matrix cosine similarity kernel. Owing to the linearity of the kernel, support vectors are combined into a rigid tensor detector for fast and efficient detection.
*Original picture (available online, 3rd September, 2014) is owned by Tomasz Dunn, and licensed under Creative Commons Attribution 2.0 Generic (CC BY 2.0)
Laplacian Object: computing a query subspace that preserves intrinsic image geometry — on left, the proposed two-layer hierarchical model is shown where top layer of global context (in the form of an affinity graph) guides the bottom up aggregation of local information from low level descriptors. On right, locality preserving projection with the graph Laplacian is used as a mathematical framework to represent the two-layer hierarchy.
|
| ||
 |
 |
Example detections on UIUC car test set are shown here. Single scale car detection (the query image is shown top-left), and Multiscale car detection (the same query image as used in single scale experiment is used here). A red bounding box indicates highest resemblance to query image, and for other colors the colormap shown right depicts relative resemblance.
|
| ||
 |
 |
Single Scale: Example detections along with scale estimation are shown in the figure above using face as query (bottom left) to detect faces in MIT-CMU face data set. The correct bounding box results from the maximum likelihood estimate of probable set of scales. Multi Scale: Sample detections along with pose estimation for the most general case is shown above with MIT-CMU face images. The appropriate scale as well as orientation of the query both result from the maximum likelihood estimates over the set of probable scale and orientations.
Charade (1963): in leftmost column the user defined query object is highlighted, and example detections are displayed on right.
Dressed to kill (1980): in top row, leftmost column, the user selects the bow-tie as query object, and sample detections are shown in right panels. In second row, leftmost column, the selected biscuit jar as query gets detected in subsequent frames in the middle of heavy clutter, scale change and partial occlusion.
Ferris Buller's Day off (1986): in top row leftmost column, the user selects the wall painting (within camera focus), and subsequent detections include cases with heavy out-of-focus instances and partial occlusions. In second row, the selected jersey number is detected against considerable scale change and minor off-the-plane distortion. Lastly, in the third and fourth rows, we see the red wing logo detected correctly on the T-Shirt despite some challenging distortions like scale and varying aspect ratio.
He, Xiaofei, Shuicheng Yan, Yuxiao Hu, Partha Niyogi, and Hong-Jiang Zhang, <b>Face recognition using Laplacianfaces, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2005
Eli Shechtman and Michal Irani, Matching local self-similarities across images and videos, IEEE Conference on Computer Vision and Pattern Recognition, 2007
Hae Jong Seo and Peyman Milanfar, <b>Training-free, generic object detection using locally adaptive regression kernels, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2010
Dubout, Charles, and François Fleuret, <b>Exact acceleration of linear object detectors</b>, European Conference on Computer Vision, 2012