http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
Model Diet: A Simple yet Effective Model Compression for Vision Tasks
Jongmin Lee,Armagan Elibol,Nak Young Chong 제어로봇시스템학회 2021 제어로봇시스템학회 국제학술대회 논문집 Vol.2021 No.10
Computer vision coupled with machine learning algorithms has greatly helped mobile robotic platforms become more intelligent and capable of performing in the real world. Specifically, Convolutional Neural Networks (CNNs) have achieved a high accuracy on a range of visual perception tasks (e.g., object detection, classification, segmentation, and similar others). One of the bottlenecks in CNNs is their high computational requirement. This makes most of them not easily deployable on robotic platforms, since their on-board computational power is limited. Recently, Involution successfully reduced the number of parameters of CNNs by replacing all the 3 × 3 convolution kernels with involution kernels, which use 1 × 1 convolution for the kernel generation. Filter pruning methods have also successively reduced the number of parameters in CNNs. Notably, however, Involution has reshaping layers and the kernel size is unknown when loading the pre-trained model. In this paper, we propose a pruning method named Model Diet that can be applied to Involution and other CNNs. We present experimental results showing that it has better results compared with randomly initialized weights.
Semantic Mapping Based on Image Feature Fusion in Indoor Environments
Cong Jin,Armagan Elibol,Pengfei Zhu,Nak Young Chong 제어로봇시스템학회 2021 제어로봇시스템학회 국제학술대회 논문집 Vol.2021 No.10
It is of the utmost importance for the robot to understand human semantic instructions in human-robot interaction. Combining semantic information with SLAM-based maps leads to a semantic map. Deep neural networks are able to extract useful information from the robot’s visual information. In this paper, we integrate the RGB feature information extracted by the classification network and the detection network to improve the robot’s scene recognition ability and make the acquired semantic information more accurate. The image segmentation algorithm labels the areas of interest in the metric map. Furthermore, the fusion algorithm is incorporated to obtain the semantic information of each area, and the detection algorithm recognizes the key objects in the area. We have demonstrated an efficient combination of semantic information with the occupancy grid map toward accurate semantic mapping.
Zeyu Ding,Armagan Elibol,Nak Young Chong 제어로봇시스템학회 2022 제어로봇시스템학회 국제학술대회 논문집 Vol.2022 No.11
Multi-turn dialogue is the major manifestation of a conversation. Compared with single-turn dialogue, response selection is more complex as the context varies. We stress the importance of dialogue history and apply the pre-trained model BERT to assign proper weight to each utterance of a dialogue. Previous works take all the dialogue history as context to measure the matching degree of a context-response pair, causing the quadratic computational cost and truncation of longer sequences exceeding the length limitation of BERT. We propose a sentence-based method to deal with the aforementioned problems, obtaining the sentence embedding of a single unit utterance of dialogue and forming a classification token of a context-response pair. We discuss how to obtain a sentence embedding with high quality and to design the input representations in response selection. The results show that the average of the first-last layer output exhibits the best performance for obtaining a sentence representation. The proposed method, concatenating the sentence embeddings of context with the token embeddings of response candidates, is nearly on a par with the token embedding based SOTA method. Notably, the processable length of dialogue history is enlarged about ten times with a low computational cost, potentially reducing chatbot response time and inspiring user engagement.
Performance Enhancement Step for Motion Estimation via Feature-based Image Matching
Keita Miyaura,Armagan Elibol,Nak Young Chong 제어로봇시스템학회 2022 제어로봇시스템학회 국제학술대회 논문집 Vol.2022 No.11
Most of the complicated and sophisticated tasks in visual robotics applications usually build upon the image matching step as matching images of the same scene can provide important information (e.g., camera motion). Image matching is generally done via extracting and matching some distinctive points via their feature vectors. This procedure generates some mismatched points due to imperfections. Mismatched points are called outliers and identified via probabilistic methods. Since the probabilistic methods work iteratively, they generally occupy a large portion of the computational cost of the whole image matching pipeline. In this paper, we present a simple yet efficient algorithm that is employed for eliminating the outliers aiming at reducing the total number of iterations needed in the probabilistic methods. Our method is motivated by the common way of visualizing the established matches among images. We tile images together and search for parallel lines connecting correspondences. We present extensive computational and comparative experiments using both simulated data involving along with real images and using a real dataset.
A Novel Room Categorization Approach to Semantic Localization for Domestic Service Robots
Felix Yustian Setiono,Armagan Elibol,Nak Young Chong 제어로봇시스템학회 2021 제어로봇시스템학회 국제학술대회 논문집 Vol.2021 No.10
Recently, room categorization as part of indoor robot localization has become a vital topic for semantic mapping. One approach is implemented via scene understanding by integrating available object information in the scene. In this paper, a novel room association approach is proposed based on the prior knowledge of the object appearance frequency in the specific room category inside the house. The front interface of the proposed technique employs a state-of-the-art YOLOv2-based object detection framework. Detected objects and their prior appearance frequency information form the input to the proposed room association through a novel scoring approach. This scoring function avoids any limit on the number of detected objects and is capable of operating with a low object detection confidence level. The experimental results of the novel proposed technique show significant improvement over the previously developed room categorization approach. On average, the correctness score increased up to 0:8387 while the indecisiveness level of the object detection framework decreases.