mm117 Frgs Review Workshop v3
mm117 Frgs Review Workshop v3
Workshop
Mohammad Faizal Ahmad Fauzi
23rd January 2020
Multimedia University
Speaker Background
• B. Eng. (Imperial College London, UK) 1997 - 1999
• PhD (University of Southampton, UK) 2000 - 2004
faizal1@mmu.edu.my 2
Outline
FRGS Guideline 2020
FRGS Review Process
MyGrants Application
Example of Reviewers’ Comments
Q&A
FRGS Guideline 2020
Application
• Pemohon amatlah digalakkan untuk berkolaborasi dengan pihak
Industri/Agensi yang berkaitan bagi melaksanakan projek penyelidikan
tersebut (bukti dokumen adalah sekurang-kurangnya surat niat (letter of
intent) dan lain-lain dokumen yang berkaitan).
• Pemohon amatlah digalakkan untuk mengemukakan bukti carian paten
(patent search) bagi setiap permohonan projek.
• Setiap pemohon dibenarkan untuk mengemukakan hanya satu permohonan
projek sahaja pada setiap fasa.
• Permohonan baharu bagi mereka yang telah diluluskan FRGS pada fasa yang
terdahulu akan diberi pertimbangan dengan bukti projek-projek tersebut telah
75% selesai di dalam Modul Pemantauan Sistem MyGRANTS dan disahkan
oleh RMC Institusi. Penyelidik dipohon untuk mengemaskini prestasi projek
penyelidikan secara atas talian melalui Sistem MyGRANTS pada setiap Februari
dan Ogos setiap tahun.
• Tempoh perkhidmatan Ketua Penyelidik mestilah berbaki sekurangkurangnya
dua (2) tahun di Institusi masing-masing mulai daripada tarikh tutup
permohonan FRGS.
National Priority Areas
Setiap permohonan hendaklah memenuhi salah satu Sektor Keutamaan Negara
(National Priority Areas - NPAs) seperti berikut:-
*Nota: Sektor Keutamaan ini adalah merangkumi IR4.0 (Cyber Security, Big Data,
Artificial Intelligence) yang merujuk kepada Dasar Negara Mengenai Industri 4.0
(National Policy on Industry 4.0).
Research Clusters
Setiap permohonan hendaklah memenuhi salah satu kluster
penyelidikan seperti berikut:-
This project aims discover novel methods for automatic spotting and recognition of
Objectives spontaneous facial micro-expressions. Firstly, this project endeavors to create the first-ever
dataset for spontaneous micro-expressions elicited from naturalistic situations to challenge
this growing research field. Existing datasets were constructed using only emotional video
Research clips as visual stimuli. Secondly, the crucial first task of micro-expression spotting demands for
novel algorithms that can effectively locate the onset, offset and apex frames from a
methodology continuous video to a good level of precision. Current heuristic-based methods for measuring
their occurrences do not generalize well across different persons. Finally, present state-of-the-
art recognition performance can be improved with new spatio-temporal feature
representations that can robustly amplify and discriminate between micro-expression signals
from different categories.
The expected output of this project is three-fold in terms of new fundamental knowledge and
Expected outcomes findings: Algorithms for representing and spotting micro-expression occurrences, theoretical
spatio-temporal models for better recognition of micro-expressions, and valuable know-how
into the construction of spontaneous micro-expression dataset from naturalistic scenarios.
This project will have far-reaching impacts towards the well-being of the society and public
Significance of safety of the nation with its strong relevancy to governmental policies.
output
Research Background
Research Background
Problem Statement
Example
A micro-expression is a brief and involuntary facial movement which reveals a genuine emotion that a person
tries to conceal [1, 2]. The duration of a micro-expression is in the range of 1/3 to 1/25 seconds (the generally
acceptable upper limit is 1/2 second at most [2] ), and it typically appears with low intensity, almost difficult to
be observed with the naked eye. As such, a person exhibiting micro-expressions is usually unaware or unable to
control it spontaneously through willpower. While psychologists have been studying facial micro-expressions
since the 1960’s, only until the last few years, it has begun to garner more attention in affective computing
studies. Machine recognition of general facial expressions has been intensively studied in the field of computer
vision for decades [3], but very little research has been done to process and analyze micro-expressions using
computers. The analysis of spontaneous micro-expressions has many potential applications; particularly for
clinical diagnosis of the emotional well-being of the patient, and various mental/psychological conditions such as
autism and depression; and for criminal interrogation, where micro-expressions formed essential clues for
detecting lies and cheating among suspects.
Analysis of spontaneous facial micro-expressions is very challenging, even for human beings [1], due to their
short duration and subtleness in intensity. Generally, given a continuous video clip of a person-of-interest,
analysis of micro-expressions includes performing two inter-connected tasks: “spotting”, followed by
“recognition”. A recent work collecting spontaneous micro-expressions [4] pointed out that human annotators
found the task of spotting challenging and heeded the advice of Ekman to first view the video frame-by-frame
and then with increasing speed [5]. Ekman’s study [6] shows that people without training were only able to
perform ME recognition slightly better than chance, while a trained psychologist is able to perform better but
only up to an accuracy of 40+%. This was not a considerable obstacle for detecting and recognizing normal or
“macro” expressions. Hence, there are several challenges in this relatively new field of work that is worth
pursuing, particularly on the state of current data, and the dual tasks of ME spotting and recognition.
Problem Statement
Example
Elicitation of MEs is difficult because it is only present at specific situations, when a person tries to suppress felt
emotions but fails to do so, resulting in a leaked emotion manifested as micro-expressions [2]. Both current state-of-the-
art datasets, i.e. SMIC [4] and CASME II [7] (CASME [8] is the predecessor of CASME II), were collected by asking subjects
to watch emotional video stimuli while attempting to keep neutralized faces within a “high-stakes” situation (they will be
penalized for any emotion leaks). However, facial MEs in more naturalistic situations should not be limited to this
particular paradigm of elicitation. Realistic scenarios such as face-to-face conversations/interviews, or a high-stake
situation involving telling of lies about what they have read or seen previously, may also exhibit MEs. Moreover, these
situations reside in a more unconstrained environment (than in a very constrained studio setting) that requires dealing
with other irrelevant facial movements. Also, both the SMIC and CASME II datasets were constructed by extracting short
video samples containing a single emotion each. This is inadequate to cater for a complete spot-then-recognize
procedure, as naturalistic situations may typically comprise of a number of micro-expressions, among a spate of other
kinds of movements.
In order to recognize what type of MEs are present, the MEs must first be spotted, or detected in the video. Thus, the
capability of recognition is strongly dependent on the success of correctly spotting the MEs. Manual spotting of the onset
(starting), apex (strongest) and offset (ending) frames in video is often a time-consuming task that requires high
reliability. Compared to methods for recognition, spotting is also a significantly understudied problem in this area.
Existing approaches are fraught with many impediments, i.e. optimal threshold parameters are overly sensitive and hard
to identify [9], the presence of eye-blinks which had to be manually removed [10]. This is clear indication that research in
micro-expression spotting is still at an early stage.
While the original recognition baselines of CASME II and SMIC starting off (in 2014) at an F1-score of 0.35 and 0.40
respectively, the latest state-of-the-art methods [11] and [12] (our group’s latest work) are now achieving a top score of
0.57 and 0.60 for CASME II and SMIC respectively. Nevertheless, this is still far from being reliable enough for use in
automated recognition systems. Interestingly, many methods in literature have worked primarily on two separate aspects
of the recognition pipeline: temporal pre-processing [12, 13] and spatio-temporal feature representation [14, 15] without
exploiting how they can be seamlessly integrated. On ME data, traditional feature representations [16] do not fare as
well as with visual data from other domains such as normal facial expression and human actions/activities. It is likely that
subtle facial changes across time are difficult to capture and encode into meaningfully discriminative information.
Moreover, the use of motion magnification has not been thoroughly explored; preliminary attempts were made [17, 18]
but improvements had been limited.
Hypothesis
Example
Due to the above-mentioned problems and challenges, we hypothesize that facial micro-expressions will be a fast-
growing field of study in computer vision and affective computing in the coming years. Research on normal facial
expressions have reached a saturated stage with off-the-shelf products such as Affectiva [19] and Emotient [20]
able to attain almost perfect emotion/expression recognition rates. Meanwhile, research on subtle micro-
expressions remains a new challenge. As of now, there are only three research groups in the world who have been
actively working in this new area since 2014 – namely Oulu University, the Chinese Academy of Sciences (CAS) and
Multimedia University (MMU) (our group).
The recent emergence in advanced techniques such as efficient dense optical flow [21] and Eulerian motion
magnification [22] presents viable solutions for subtly changing facial motions. New spatio-temporal encoding
methods [23] that are robust against noise can be potentially explored and innovated further for this domain of
work. It is envisaged that machine-assisted analysis of facial micro-expressions will in the foreseeable future,
provide auxiliary support for social and scientific domains such as forensic science, criminal investigation, medical
diagnosis and psychotherapy.
Research Questions
Example
The direction of this research can be summarized with the following questions:
With the fundamental nature of the proposed project, the expected research outcomes have deeply-rooted
concepts that are directly transferable to various applied domains, and have far-reaching impacts on the well-
being of the society and public safety of the nation. The prediction of human emotional states, especially in
situations where patients attempt to withhold their emotions is particularly beneficial for clinical assessment of
psychological conditions such as autism, schizophrenia and depression. This is in line with Malaysian government's
policy under the Healthcare National Key Economic Areas (NKEA), which aims to create a supportive ecosystem for
clinical research and diagnosis.
In the aspect of criminal investigation and forensic science, our contributions enable investigators to see through
the hidden emotions of deceptive suspects and criminals, who may attempt to conceal their actual agenda and
motives. Hence, this project has strong relevancy in terms of its potential contribution towards supporting the
National Key Results Areas (NKRA) under Governmental Transformation Policy (GTP) 2.0 in two priority areas: (I)
Fighting Corruption and (II) Reducing Crime. Interrogations of suspects conducted by police or other authorities,
and drug addict rehabilitation efforts can be greatly enhanced with the help of such technology.
Objectives
Objectives
Objectives
Example
This research study embarks on the following objectives:
1. A Good Fit
59
Other Comments
• A new idea is claimed but insufficient technical details of the idea
are given for the committee to be able to judge whether it looks
promising.
– Since the committee cannot be expert in all areas there is a danger
of overwhelming them with technical details, but it is better to err by
overwhelming them than by underwhelming them. They will usually
get an expert referee to evaluate your idea.
• The proposers seem unaware of related research.
– Related work must be mentioned, if only to be dismissed. Otherwise,
the committee will think that the proposers are ignorant and,
therefore, not the best group to fund. The case for support should
have a list of references like any paper, and you should look at it to
check it has a balanced feel - your referee will do so. Do not make
the mistake of giving references only to your own work!
60
Other Comments
• The proposal is badly presented, or incomprehensible to all but
an expert in the field.
– Remember that your proposal will be read by non-experts as
well as (hopefully) experts. A good proposal is simultaneously
comprehensible to non-experts, while also convincing experts
that you know your subject. Keep highly-technical material in
well-signposted section(s); avoid it in the introduction.
• The proposers seem to be attempting too much for the funding
requested and time-scale envisaged.
– Such lack of realism may reflect a poor understanding of the
problem or poor research methodology.
• The proposal is too expensive for the probable gain.
– If it is easy to see how to cut the request for
people/equipment/travel, etc. to something more reasonable
then it might be awarded in reduced form. More likely, it will
be rejected.
61
Common Failings
1. Did not follow application guidelines exactly.
2. Not new or lack of original ideas
- e.g. routine application of known techniques.
3. Lack of knowledge of published relevant work
4. Diffuse, superficial or unfocused research plan
5. Lack of experience in the essential methodology
- uncritical approach
6. Insufficient experimental detail
- or questionable reasoning in experimental approach
7. Unrealistically large amount of work
8. Unacceptable scientific rationale
9. Uncertainty concerning the future directions
62
Messages From RMC
Source: MOHE 2018
Budget
FRGS Guidelines
Vote 11000 Labour GRA (Pursuing Master: RM1800,
Pursuing PhD: RM2300)
Vote 21000 T & T Conference & training: 20%
Total T & T: 40% of project cost
Vote 24000 Rental Relates to research