Introduction:
Time: 40 minutes, slides 9 am – 9.40 am
Objective: In this first part of the tutorial we introduce the basic scope of person re-identification, the nature of the problem, issues, datasets, modalities and why this approach became so popular lately.
The introduction to the topic is broke down in the following items:
- Definition of the person re-identification problem;
- Why person re-identification is useful [4];
- Use cases:
- Re-acquire people in camera networks;
- Recognize people in biometric systems;
- Automatic profiling of customers/visitors/tourists;
- Introduction of the basic scenarios, modalities [2] and probe/gallery paradigm.
- Introduction of the main datasets (VIPeR, CAVIAR4REID, iLIDS, ETHZ, 3DPeS, CMV100, etc.) used for algorithm evaluation [4]
- Presentation of the standard evaluation tools (Cumulative Match Characteristic, CMC) [4] and clarifications about general vocabulary (recognition rate at first rank, etc.).
Person Representation
Time: 40 minutes, slides 9.40 am – 10.20 am
Objective: In this part we present techniques that try to build a discriminative representation of a person while preserving viewpoint and illumination invariance.
This topic is organized as follows:
- How to describe an image containing a person.
- Which cues are important?
- Colors
- Texture
- Saliency
- Is background/foreground segmentation important?
- Is part segmentation important?
- Overview of main representation milestones:
- SDALF descriptor [1]
- WHOS descriptor [2]
- Other descriptors
- Short, practical MATLAB session on how to compute common descriptors given an image
Matching People across Views
Time: 40 minutes, slides 10.20 am – 11.00 am
Objective: In this part of the tutorial we address how to match different people considering the probe/gallery paradigm and how to better exploit the gallery images that describe a target. We then introduce different matching strategies from a baseline (Nearest Neighbor) matcher to more complex statistical models that try to solve the gap of matching people across non-overlapping cameras.
This topic is organized as follows:
- Recap on Modalities and probe/gallery paradigm
- Nearest Neighbor (NN)
- NN with combined distances (SDALF) [1]
- ISR (Sparse Iterative Ranking) method [2]
- Metric Learning methods
- KCCA-based (Kernel Canonical Correlation Analysis) method [2, 3]
- Short, practical MATLAB session of matching methods.
Coffe Break
Time: 10 min , 11.00 am – 11.10 am
Practical Code Session
Time: 25 min, code session, 11.10 am – 11.35 am
Objective: In this part of the tutorial we show and run some didactical MATLAB tool to perform some re-identification tasks:
- Person Description
- Matching
- Method Evaluation
A case of study, Mnemosyne project
Time: 25 min, slides 11.35 am – 12.00 pm
Objective: In this final part, we discuss a real application of person re-identification. The system is used to profile tourist interests while they are walking in a museum in order to improve their experience about the artwork. We discuss solved problems along with open issues and we introduce a new dataset acquired in this scenario.
This topic closes the tutorial as follows:
- Introduction of the Mnemosyne project: scope, issues and solved problems [5]
- Show some results and a video demo.
- Discuss open issues in real-world applications.
- Wrap-up and conclusions
References:
- M. Farenzena, L. Bazzani, A. Perina, V. Murino, and M. Cristani, “Person re-identification by symmetry-driven accumulation of local features,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit.,
2010, pp. 2360–2367. - Lisanti, G.; Masi, I.; Bagdanov, A.D.; Bimbo, A.D., “Person Re-Identification by Iterative Re-Weighted Sparse Ranking,” Pattern Analysis and Machine Intelligence, IEEE Transactions on , vol.37, no.8, pp.1629,1642, Aug. 1 2015
doi: 10.1109/TPAMI.2014.2369055 - Giuseppe Lisanti , Iacopo Masi , Alberto Del Bimbo, Matching People across Camera Views using Kernel Canonical Correlation Analysis”, Eighth ACM/IEEE International Conference on Distributed Smart Cameras, 2014.
- D. Gray, S. Brennan, and H. Tao, “Evaluating appearance models
for recognition, reacquisition, and tracking,” in Proc. 10th IEEE
Workshop Perform. Eval. Tracking Surveillance, 2007, pp. 41–48. - Federico Bartoli, Giuseppe Lisanti, Lorenzo Seidenari, Svebor Karaman, Alberto Del Bimbo,”MuseumVisitors: a dataset for pedestrian and group detection, gaze estimation and behavior understanding”, Proc. of CVPR Int’l. Workshop on Int.’l Workshop on Group And Crowd Behavior Analysis And Understanding – 2015