How can we help you? Search AIC
SK telecom’s AI Center participated in ICML and CVPR, two major conferences in the field of Machine Learning and Computer Vision, and installed company booths and held the recruiting events from June 10th to 20th. In order to identify the latest trends, researchers from T-Brain, a research group within the AI Center, presented Tutorial on recent research, Workshop on individual topics, and Main Conference and Keynote Speech as well as participated in networking session with conference attendees from around the world.
During CVPR, T-Brain presented a total of three Workshop papers and once again proved its technologies in the field of AI.
First, T-Brain participated in Davis Challenge on Video Object Segmentation and achieved the 3rd place. Davis Challenge, which is an abbreviation of Densely Annotated Video Segmentation, is a Video Object Segmentation challenge jointly held with CVPR. Under the given dataset, participants have to compete for how to perceive and segment objects employing one of the three techniques - Semi-Supervised Learning, Interactive or Unsupervised Learning - and achieve SOTA (State of the Art) performance. From T-Brain, three researchers participated in Unsupervised Learning Track and the relevant research on Key Instance Selection for Unsupervised Video Object Segmentation was presented during the Workshop of Davis Challenge on Video Object Segmentation. This research assumes that the number of objects perceived by a human from the video is limited and the objects are also perceived in initial frames. Based on this, we continuously measure and track similarities of the objects detected within the following frames.
In the Workshop on Mutual Benefits of Cognitive and Computer Vision, a research on Bilinear Attention Networks for Visual Grounding was presented. This research introduces Visual Grounding using Bilinear Attention Networks (BAN) to find the main objects in the given image and when the sentence is given, it deduces the relationship between objects from the image and words/syllables from the sentence. Through this research, we can identify that while human’s sentence comprehension may affect the visual reasoning and vice versa, BAN can compute this interaction at once.
Finally, in the Workshop on Visual Question Answering and Dialog, a research on Relational Bilinear Attention Networks with the Assumption of Linear Compositionality was presented. In this paper, we used BAN to confirm the linear compositionality, which is frequently appeared in word vectors, can be applied for image representation vector. For example, word vectors that represent the country, such as ‘Germany’ or ‘China’ and word vectors that represent the capital of the country, such as ‘Berlin’ and ‘Beijing,’ can be explained with the relational vectors, which are parallel as well as with the same length. Thus, given a word vector representing an arbitrary country such as ‘France,’ this relational vector can be used to deduce a word vector ‘Paris,’ and we call this linear composability. In many of the recent researches, they frequently employed object detectors to find the visual vectors; this time, we demonstrated that this linear compositionality can be observed among visual vectors extracted by object detectors without supervised learning. For example, if you add ‘wearing’ relational vector to a ‘man’ visual vector, you can find the visual vector for the ‘shirt’ that ‘man’ is wearing as the nearest neighbor.
During the conferences, AI Center booth was set up to introduce SK telecom’s AI Center, in-house AI services and detailed research at T-Brain. Furthermore, T-Brain held four recruiting events to exchange up-to-date information and engage with participating professors, students and researchers who proceed on-going research in related fields. Kim Jiwon, a lead of T-Brain, mentioned “Many researchers who attended the conference are important individuals and will lead the development of AI in Korea. We will continuously lead the community for AI development through various research opportunities in the future.”