Various camera types, from linear perspective to wide field of view (FoV), panoramic, and hemispherical, are utilized in computer vision due to their distinct features. Images obtained from a perspective camera, which mimics the human eye are dominant in benchmark datasets such as ImageNet and MS COCO. Others obtained from wide FoV cameras are present in surveillance and person detection datasets, and autonomous driving. The proliferation of visual data acquisition through cameras with diverse FoVs has posed a significant challenge to the development of robust object recognition models. This heterogeneity necessitates the creation of an object recognition engine that can seamlessly adapt to the complexities introduced by different viewpoints, distortions, and angle variations. In this work, we leverage data augmentation to enhance the robustness of the YOLOv7 model when confronted with various camera types, such as perspective variations, distortions, and diverse viewing angles in images. We created the RMFV365 dataset, comprising 5.1 million images that encompass a broad spectrum of perspectives and distortions. This dataset serves as a valuable resource for training the model in generalized object recognition tasks. For testing purposes, we modified COHI, a benchmark fisheye testing dataset to COHI-365. The YOLOv7 model trained on RMFV365 (YOLOv7-T2) outperforms the YOLOv7 model trained on Objects365 (YOLOv7-Base), achieving a 4.8% improvement as measured by the mAP50 metric when tested on COHI-365. Also, YOLOv7-T2 achieves a close result as YOLOv7-Base when tested on the original Objects365 data.

In industrial applications, Human Pose Estimation (HPE) is crucial for enhancing both automation and human-computer interaction. This study investigates the impact of synthetic data on HPE model efficacy, particularly examining the performance of the YOLOv8 algorithm. Using Nvidia Omniverse Isaac Sim, we created a synthetic dataset called ISAAC, designed for various complex scenarios. This tool was chosen for its ability to simulate highly realistic and intricate industrial contexts with advanced physics and AI capabilities. The inclusion of this synthetic dataset significantly enhances model accuracy, evidenced by up to a 19% increase in mean Average Precision (mAP) at an Intersection over Union (IoU) of 0.5, and a 12% improvement across the 0.5-0.95 IoU range compared to traditional datasets. These results highlight the substantial advantages of synthetic data in training more accurate and robust HPE models, advocating for the integration of innovative data solutions in the field of computer vision.

Object recognition is a machine learning problem that involves the correct classification and localization of objects in an image. Object recognition has found wide applications in Industry 4.0, surveillance, and autonomous driving. Prominent datasets such as Pascal VOC, ImageNet, and MS COCO have further spurred advances in object recognition. However, the images in these datasets were captured using ordinary perspective lenses. Unlike ordinary cameras, fisheye or hemispherical cameras have a wide FoV reaching 180°. Due to this property, they can acquire wide panoramic images, thus, capturing more objects in a scene as compared to ordinary perspective lenses. In this work, we provide the first benchmark testing dataset of real fisheye images containing commonly-occurring objects in the human environment obtained with hemispherical cameras. The dataset comprises 1,000 images containing 39 classes with a total number of 14,218 object instances. We trained the YOLOv7 model on the original MS COCO and transformed FisheyeCOCO datasets and tested the obtained models with our dataset. The results indicate that training the model on the combination of original and transformed datasets improves the performance by 2.5% compared to using these datasets separately.

Social robots are increasingly being used as a medi-ator in robot-assisted autism therapy to improve children's social and cognitive skills. Engagement is one of the key measurements used to evaluate the therapeutic interventions' effect on children. While “engagement” is broadly used, it has been challenging to find a consensus about its definition in the community. With this paper, we explore the use of a data-driven approach to investigate the extent to which a model on engagement built on one dataset transfers to another. We utilized two publicly available datasets of engagement recognition, namely PInSoRo and Qamqor datasets, with an attempt to achieve a higher accuracy taking into account the transferred knowledge. The accuracy of 83.18% was obtained on the PInSoRo dataset of child-child interactions with face and body keypoints. We have used the methodology of transfer learning to improve the classification accuracy on the Qamqor dataset. The best result obtained is a 71.89% accuracy on the Qamqor dataset. This suggests that more data with similar keypoints is needed to achieve better accuracy when utilizing transfer learning from one dataset to another dataset.