Multi-Rater Consensus Learning for Modeling Multiple Sparse Ratings of Affective Behaviour

IEEE Transactions on Affective Computing, Journal Year: 2023, Volume and Issue: 15(3), P. 859 - 871

Published: July 20, 2023

The use of multiple raters to label datasets is an established practice in affective computing. principal goal reduce unwanted subjective bias the labelling process. Unfortunately, this leads key problem identifying a ground truth for training affect recognition system. This becomes more relevant sparsely-crossed annotation where each rater only labels portion full dataset ensure manageable workload per rater. In paper, we introduce Multi-Rater Consensus Learning (MRCL) method which learns representative model that accounts rater's agreement with other raters. MRCL combines multitask learning (MTL) regularizer and consensus loss. Unlike standard MTL, approach allows learn predict while explicitly accounting among We evaluated our on two different based spontaneous body movement expressions pain behaviour detection laughter type respectively. naturalistic were chosen forms (different affect, observation stimuli, raters) they together offer evaluating approach. Empirical results demonstrate effective modelling from multi-rater annotation.

Language: Английский

An MQTT-Based Student Condition Monitoring System for Physical Education

Published: July 1, 2023

With the development of Internet Things (IoT) technology, wearable devices have been widely used in different fields. However, few studies focused on application and IoT technologies student movement detection systems for Physical Education (PE). This paper mainly designs an Message Queuing Telemetry Transport (MQTT)-based condition monitoring system physical status monitoring, where network transmission from perception layer to a multi-user scenario is considered. The proposed consists data acquisition module, communication module message transmission, analysis application. (MQTT) protocol, as low-overhead low-bandwidth-consumption instant messaging applied enable be published clients that has subscribed corresponding topics real-time. During experiment, each client publishes amounts broker, simulating multiple users sending it sends received through flow function. results show MQTT protocol low latency situations. Also, able provide real-time reliable services with minimal volume limited bandwidth, which reveals feasibility its smart education.

Language: Английский

Movement Representation Learning for Pain Level Classification

IEEE Transactions on Affective Computing, Journal Year: 2023, Volume and Issue: 15(3), P. 1303 - 1314

Published: Nov. 20, 2023

Self-supervised learning has shown value for uncovering informative movement features human activity recognition. However, there been minimal exploration of this approach affect recognition where availability large labelled datasets is particularly limited. In paper, we propose a P-STEMR (Parallel Space-Time Encoding Movement Representation) architecture with the aim addressing gap and specifically leveraging higher pain-level classification. We evaluated analyzed using three different across four sets experiments. found statistically significant increase in average F1 score to 0.84 pain level classification two classes based on compared use hand-crafted features. This suggests that it capable representations transferring these from data captured lab settings levels messier real-world data. further efficacy transfer between can be undermined by dissimilarities population groups due impairments behaviour motion primitives (e.g. rotation versus flexion). Future work should investigate how effect differences could minimized so healthy people more valuable learning.

Language: Английский