EVEX - Object Evidence Extraction
Evidences = image features = keypoints = interest points
Local descriptors = interest point descriptors
The goal of this project is to study existing and develop new methods for accurate detection of local image features aka object evidences. The detection of local image features is needed in the recently emerged object detection, localisation and recognition approach where image features (local descriptors) are separately detected and then "hypotheses" of object location and pose are searched based on proper spatial configuration (constellation model) of the features.
This project is a sub-project of more general project, OBJECT - Object Detection and Recognition, and the results will be utilized in the main project.The current main objectives are:
- State-of-the-art method for extraction of local image features (supervised approach)
- Multiresolution Gabor feature based local descriptor
- Descriptor ranking and classification based on the Gaussian mixture model
- FACEDETECT - Face Detection
The proposed state-of-the-art image feature extraction method was accepted to be published in IEEE-TIP and the results will be published in the PhD thesis of Jarmo Ilonen.
|Joni Kämäräinen||WWW||Project leader, lecturer|
|Jarmo Ilonen||WWW||Post-graduate researcher|
|Pekka Paalanen||WWW||Post-graduate researcher|
Data setsThe following publicly available data sets have been used in our experiments:
- The Extended M2VTS Database (XM2VTS) (only colour still images used) (by CVSSP, University of Surrey, UK).
- Training face landmarks: XM2TrClC1/
- Evaluation face landmarks: not used
- Testing face landmarks (d_eye used): XM2TeImC1/
- XM2VTS/non-frontal (mpeg7) ground truth annotated by Joni:
- For developing spatial constellation search methods you may find useful our 100 best landmark candidate of each type (automatically extracted as reported in our IEEE-TIP2007 article):
- The Banca Database (by CVSSP, University of Surrey, UK).
|Researchers working on local descriptors or other similar methods|
|Dr Cordelia Schmid (INRIA)|
|Prof David Lowe (University of British Columbia)|
|Dr Timor Kadir (University of Oxford)|
Do not hesitate to to contact authors (e-mail, etc.) in order to retrieve copies or reprints of the following publications.