This paper addresses the task of fine-grained label learning in object detection with the weak supervision of auxiliary information attached to images. Most of the recent work focused on the label... Show moreThis paper addresses the task of fine-grained label learning in object detection with the weak supervision of auxiliary information attached to images. Most of the recent work focused on the label prediction for objects in the same category space as in training data under the fully-supervised learning framework and cannot be expanded to the learning of more fine-grained categories that have not been defined in training sets. In this paper, we propose a new weakly-supervised learning approach, called label inference curriculum network (LICN), to detecting objects and learning their fine-grained category labels based on supervision of captions via curriculum learning. First, we build a semantic mapping based on embedding techniques and a knowledge base to measure the correspondence between coarse labels and fine-grained label proposals; second, we introduce a label inference curriculum network, which ranks the order of training samples by the complexity of samples. We construct two datasets, namely FG-COCO and FGs-COCO, consisting of both coarse and fine-grained labels based on MS COCO and Visual Genome to train and test our approach. Experimental results demonstrate the effectiveness of our proposed LICN model, and LICN-E2C achieves an improvement of 1.7% mAP with 0.5:0.05:0.95 IoU compared with the LICN-C2E on the FG-sCOCO test dataset. Show less
Beurden, M. van; Brouwer, A.M.; Baardewijk, J.U. van; Binsch, O.; Vermetten, E.; Roijendijk, L. 2020
Feedback of physiological responses have a great potential to support virtual training paradigms aimed to increase cognitive task performance under stressful threatening conditions. In the current... Show moreFeedback of physiological responses have a great potential to support virtual training paradigms aimed to increase cognitive task performance under stressful threatening conditions. In the current study, we examined the sensitivity of a range of physiological indicators derived from electrodermal activity (EDA), blood pressure (BP) and heart rate (HR) to measure stress as induced by the threat of an electric shock (ES). In contrast to previous work that studied physiological stress responses compared to a rest condition, we compared conditions with high cognitive load combined with stress caused by threat of an ES, to conditions with high cognitive load without such stress. Twenty-five participants performed a cognitively demanding task in an experimental setup. At certain 10 s time intervals, indicated by a continuous tone, participants were either asked to do their best and increase cognitive task performance (non-threat condition), or they were told that they could receive an ES during this interval if cognitive task performance was not high enough (threat condition). Physiological measures, task performance and self-reported measures of stress and workload were analysed. Task performance and self-reported measures of stress and workload were roughly the same in both conditions. Especially EDA measures were affected by the threat of an ES. Threat and non-threat conditions could be distinguished with an across-participant classifier using EDA and BP features with an accuracy of 70%. These results suggest that EDA and BP can be used to evaluate stress coping training paradigms or to individually adapt the stress levels in virtual training environments. Show less