We have developed methods to:
- co-organize multiple photo sequences;
- organize photo collections based on the automatically discover identities.
Concurrent Photo Sequence Organization [Dataset available]
During a social event (a birthday party, a marriage, a trip…), it is pretty common each member of the social group takes is own pictures. The photo sequences collected by the members of the social group have interesting properties:
- they are part of the same “story”: each member has captured a different “point of view” of the story;
- they have been concurrently acquired but are not temporally synchronized;
- they may be overlapping (in content) but may also be pretty different;
- they have been acquired with different devices and settings.
We tackle an interesting problem: how to co-organize these pictures based on the underlying hidden story?
 |
As all the sequences have been taken at the same social event, it is likely they would share a similar temporal-visual structure. Therefore, our goal is to learn the temporal and visual structure of the social event from a reference sequence, and then transfer it to the unprocessed sequences. |
Please, for more details, refer to our work:
Lo Presti L., La Cascia M., “Concurrent Photo Sequence Organization”, Springer Journal on Multimedia Tools and Applications (MTAP), 2012, (pdf)
The Concurrent Photo Sequence Dataset is made available!! Take a look…
People Re-Identification in Personal Photo Collection
Personal Photo Collection Organization based on “who is in the picture?” is hard because:
- face detection and recognition is “in the wild”, under varying pose and illumination conditions;
- the number of identities in the collection is a priori unknown;
- in case of “family” collection, recognition is harder!
We adopt tracking by detection methods to organize a photo collection taking advantage of the Mutual Exclusivity property of the pictures: an identity cannot appear twice in the same picture.
 |
We have used a semi-supervised approach to learn on-line the appearance model of each identity. The decisions of the trained classifiers are fused by a probabilistic framework and used to self-train the appearance models (collaborative training). |
Please, for more details, refer to
Lo Presti L., La Cascia M., “An On-line Learning Method for Face Association in Personal Photo Collection”, Elsevier Image and Vision Computing (IMAVIS), 2012, (pdf)
| We have used a joint probabilistic data association approach to organize the faces detected in a photo sequence; the set of the most probable associations between detected faces and discovered identities is the maximum matching in a bipartite graph.
Please, for more details, refer to |
 |