-
Notifications
You must be signed in to change notification settings - Fork 24
General Planning
Sub Projects | Meeting Reports
Objectives of this meeting:
- Establish the sub-projects
- Establish a High level Timeline
Group leaders: Yoshua Bengio (YB), Pascal Vincent (PV), Aaron Courville (AC), Roland Memisevic (RM), Chris Pal (CP)
Project Coordinator: Myriam Côté
Priorities are coded as:
HP: High Priority
MP: Medium Priority
LP: Low Priority
The short-term priority tasks (HP) is to establish an end-to-end baseline:
-
(HP) Code Sharing directory organization, data files organization, integration of different types of codes (e.g. with and without Pylearn2), wiki / results organization
PV (lead), Pascal Lamblin, Ian Goodfellow (consultant) -
(HP) Heuristic segmentation(from raw files to sets of face-tubes, each being a sequence of bounding-box images) currently using Picasa and simple rules
CP (lead), Mehdi -
(HP) Data organization (both of the challenge data and of auxiliary datasets) standardized access to all the relevant datasets in Pylearn2
Sub-task: preprocessing tools (aspect ratio, warp, etc.) transformations
PV (lead), Vincent, Mehdi, Pascal L. -
(HP) Performance evaluation setups (for the main tasks of interest, but mostly for the challenge measure) * some SOTA papers study needs to be done beforehand
AC (lead), PV, Pascal Lamblin -
(HP) Single-frame emotion classification (what we used to do) (many different learners could be compared there)
PV (lead), Samira, Mehdi, Ian, Heng, Salah -
(MP) Emotion classification from speech - we might get data from the 'emotion challenge' at Interspeech 2009 (need to apply) or Humaine (only 48 clips) or Vera am Mittag, see http://emotion-research.net/databases
Sub-task: emotion from speech data statistics and setup.
YB (lead), Yann, Nicolas, Razvan, (need to do this quickly because both leave in May), Guillaume A., Stephan -
(MP) Explore temporal aggregation methods simplest is max-pooling but may be fooled by short spurious peak of wrong emotion detected, MIL/noisy or could be better, and more advanced includes introducing a segmentation latent variable (reaching into the topic of integrated recognition & segm.)
YB (lead), PV, Razvan, Guillaume D., Stephan -
(MP) Integrated classication (bimodal systems), combining multiple sources of evidence, temporal change, activity recognition
AC (lead), CP, RM, Atousa, Pierre-Luc, Samira -
(LP) Integrated segmentation & recognition methods (consider segmentation and other variables like actor identity as latent variables to be marginalized or maxed-out)
AC (lead), YB, Heng, Nicolas, Sina -
(LP) Keypoint detection & tracking (may be used for face tracking, face detection, pose detection, as components of some emotion detection solutions)
AC, PV, CP, Pierre Luc, Hani (apply current eye detector to challenge dataset), Samira
10.1 Add SFEW to keypoint dataset database. -
(LP) Face & pose detection & face tracking (possibly just using the above, but not necessarily)
CP, YB, AC, Caglar, Xavier -
(LP) Keypoint-free models (project outputs from single region model and region pair model into a shared feature space and aggregate in that space the features obtained from all observed single regions and nearby region pairs; the regions need not be those surrounding keypoints nor do they need to be classified in a hard way a to their keypoint identity)
RM (lead), AC, Vincent, Guillaume D. -
(LP) (Optional) Person identification (in the sense of learning a distributed representation that puts faces of the same person near each other)
YB (lead), Salah -
(LP) Find and compare with SOTA on all these tasks, in particular AFEW and bimodal emotion recognition
AC (lead), Samira and the rest.
Gregoire and David remain to be assigned. Eventually Eric Martin, and maybe David Krueger.
Each sub-team should meet in the next few days (by April 5) to clarify what has to be done, write that down in the page of their sub-project, and get started with the work. Each sub-project leader is responsible for making sure that these meetings take place. The work planned should cover around 2 weeks of effort.