Introduction

Body gestures are an important form to reveal people’s emotions alongside facial expressions and speeches. For special occasions when people intend to control or hide their true feelings, e.g., for social etiquette or other reasons, body gestures are harder to control and thus the more revealing clues of actual feelings compared to the face and the voice. Microgestures (MG) are defined as a special category of body gestures that are indicative of humans’ emotional status. Representative instances include: scratching head, touching nose, and rubbing hands, which are not intended to be shown when communicating with others, but occur spontaneously due to e.g., felt stress or discomfort as shown in Figure 1. The MGs differ from indicative gestures which are performed on purpose for facilitating communications, e.g., using gestures to assist verbal expressions during a discussion. Although research on general body gestures is prevailing, it is largely about human’s macro body movement, lacking the finer level consideration such as the micro-gestures discussed above. As well the studies are mainly concerned with recognizing the movement performed expressively, and the link between gestures and hidden emotions is yet to be explored.

Fig. 1. Micro-gesture examples. A tennis player is talking while spontanously performing the body gestures in the post-match interview.

Motivated by the above observations, we propose to jointly hold the first MiGA challenge and workshop on micro-gesture analysis for hidden emotion understanding (MiGA) to fill the gap in the current research field. The MiGA workshop and challenge aim to promote research on developing AI methods for MG analysis toward the goal of hidden emotion understanding. As the proposed MiGA will be the first event, we propose to hold it in workshop + challenge mode with a focus on the competition part that builds and provides benchmark datasets and a fair validation platform for researchers working in the MG classification and online recognition for identity-insensitive emotion understanding. The workshop covers a wider scope than the challenge, i.e., any research that provides theoretical and practical support for gesture and micro-gesture analysis and emotion understanding.

The MiGA challenge is planned as a continuous, annual event, and each years MiGA will include different (multiple) challenge tasks, with large-scale datasets with expanding size. As it is the first year of MiGA, we will focus on two fundamental tasks of micro-gestures: classification and online recognition. The tracks of MiGA will be extended to leveraging those identity-insensitive cues to achieve hidden emotion understanding in the future. The challenge will be based on two spontaneous datasets: One is the SMG dataset published in FG2019 “Analyze Spontaneous Gestures for Emotional Stress State Recognition: A Microgesture Dataset and Analysis with Deep Learning” and the other is the iMiGUE dataset published in CVPR2021 “iMiGUE: An Identity-free Video Dataset for Micro-Gesture Understanding and Emotion Analysis”. As it is the 1st MiGA, we plan to set two fundamental challenge tasks (Tracks), and participating teams can choose to compete on one or both tasks. The challenge will be organized on the CodaLab website.

Track 1: skeleton-based micro-gesture classification from short video clips. The MG datasets were collected from in-the-wild settings. Compared to ordinary action/gesture data, MGs concern more fine-grained and subtle body movements that occur spontaneously in practical interactions. Thus, learning those fine-grained body movement patterns, handling imbalanced sample distribution of MGs, and distinguishing the high heterogeneous MG samples of interclass are the big challenges to be addressed.

Track 2: skeleton-based online micro-gesture recognition from long video sequences. Unlike any existing online action/gesture recognition datasets in which samples are well aligned/performed in the sequence, MGs samples occur spontaneously in any combinations or orders just like seen in daily communicative scenarios. Thus, the task of online micro-gesture recognition requires dealing with more complex body-movement transition patterns (e.g., co-occurrence of multiple MGs, incomplete MGs and complicated transitions between MGs, etc.) and detecting fine-grained MGs from irrelevant/context body movements, which poses new challenges that havent been considered in previous gesture research.

The rules and guidelines for the competition/challenge

1. Datasets

Datasets for the proposed challenge are available. The MiGA challenge is planned as a continuous annual event. Two benchmark datasets published on FG 2019 and CVPR 2021 are available and will be used for the challenge. The first one is the SMG dataset published in FG2019. Spontaneous Micro-Gesture (SMG) dataset consists of 3,692 samples of 17 MGs. The MG clips are annotated from 40 long video sequences (10-15 minutes) with 821,056 frames in total. The datasets were collected from 40 subjects while narrating a fake and real story to elicit the emotional states. The participants are recorded collected by Kinect resulting in four modalities, RGB, 3D skeletal joints, depth and silhouette. In this workshop, we only allow participants to use the skeleton modality. Download SMG dataset

The second dataset is iMiGUE published in CVPR2021. Micro-Gesture Understanding and Emotion analysis (iMiGUE) dataset consists of 32 MGs plus one non-MG class collected from post-match press conferences videos of famous tennis players. The dataset consists of 18,499 samples of MGs to detect negative and positive emotions. The MG clips are annotated from 359 long video sequences (0.5-26 minutes) with 3,765,600 frames in total. The dataset contains RGB modality and 2D skeletal joints collected from Open-Pose. In this workshop, we only allow participants to use the skeleton modality. Download iMiGUE dataset

Note that: 1) not all data are used for the challenge. 2) part of the data will be selected and tailored for different challenge tasks, 3) planned as a continuous event, MiGA will contain multiple parallel challenge tracks and the tasks vary in each year.

2. Evaluation

Training datasets include iMiGUE and SMG datasets while the testing dataset (without annotation released) is from iMiGUE. We deploy a cross-subject evaluation protocol by dividing the 72 subjects into a training group of 37 subjects and a testing group of 35 subjects. For MG classification track, 13,936 and 3,692 MG clips from iMiGUE and SMG279 datasets will be used for training and validating, and the remaining 4,563 MG clips from iMiGUE will be used for testing. For MG online recognition track, 252 and 40 long sequences from iMiGUE and SMG datasets will be used for training and validating, and the remaining 104 long sequences from iMiGUE will be used for testing.

MG classification track: We report Top-5 accuracy on the following subsets of the test set: 1) Overall: All segments in the test split; 2) Tail Classes: Due to the long-tailed nature of the datasets, among a total 33 classes in iMiGUE dataset, 28 classes are tail classes (approx. 57.8of the data). Since there is considerable overlap between consecutive fine-grained MG segments, we use Top-5 accuracy instead of the standard Top-1. Submissions will be ranked on the basis of Top-5 accuracy on the overall split.

MG online recognition track: We jointly evaluate the detection and classification performances of algorithms by using the F1 score measurement defined below: F1 =2*Precision*Recall/(Precision+Recall), given a long video sequence that needs to be evaluated, P recision is the fraction of correctly classified MGs among all gestures retrieved in the sequence by algorithms, while Recall (or sensitivity) is the fraction of MGs that have been correctly retrieved over the total amount of annotated MGs.

Submission format for both tracks. Participants must submit their predictions in the following format. For each sequence SequenceXXXX.zip in the data folder, participants should create a SequenceXXXX prediction.csv file with a line for each predicted gesture [GestureID, StartFrame, EndFrame] (the same format as SequenceXXXX labels.csv) (for classification track, the StartFrame and EndFrame are empty). The predictions for all the samples will be put in a single ZIP file and submitted to Codalab. The results will be evaluated on the server and displayed on the ranking list in real-time. The organization team has the right to examine the participants source code to ensure the reproducibility of the algorithms. The final results and ranking will be confirmed and announced by organizers.

Participation guidelines

Please visit the Codalab websites to join the competitions:

The 1st MiGA-IJCAI Challenge Track 1: Micro-gesture Classification

The 1st MiGA-IJCAI Challenge Track 2: Micro-gesture Online Recognition

Final results are available

We have examined all source code submitted and finalized the rankings of the MiGA challenge. Congratulations to the following competitors on their final rankings:

Track 1: Micro-gesture Classification

1. gkdx2, team name 'HFUT-VUT'

2. HHuang, team name 'NPU-Stanford'

3. ChenxiCui

Track 2: Micro-gesture Online Recognition

1. HHuang, team name 'NPU-Stanford'

2. gkdx2, team name 'HFUT-VUT'

Important Dates (might slightly adjust later)

The timeline for the Challenge will be organized as follows:
• Mar 27, 2023. Call for Challenge online. Registration starts.
• Mar 31, 2023. Release of training data.
• May 2, 2023. Release of testing data.
• May 12, 2023. Final testing data and result submission. Registration ends.
• May 15, 2023. Release of challenge results.
• June 02, 2023. Paper submission deadline (workshop).
• June 06, 2023. Notification to authors.
• June 10, 2023. Camera-ready deadline.
• August, 2023. MiGA IJCAI 2023 Workshop, Macao, China.