Published as a conference paper at ICLR 2025S ELF-SUPERVISED M ONOCULAR D EPTH E STIMATION R OBUST TO R EFLECTIVE S URFACE L EVERAGED BY T RIPLET M INING Wonhyeok Choi 1,, Kyumin Hwang 1,, Wei Peng 2, Minwoo Choi 1, Sunghoon Im 1, Electrical Engineering and Computer Science 1, Psychiatry and Behavioral Sciences 2 Daegu Gyeongbuk Institute of Science and Technology 1, Stanford University 2 South Korea 1, USA 2 {smu06117,kyumin,subminu,sunghoonim} @dgist.ac.kr 1, wepeng@stanford.edu 2 A BSTRACT Self-supervised monocular depth estimation (SSMDE) aims to predict the dense depth map of a monocular image, by learning depth from RGB image sequences, eliminating the need for ground-truth depth labels. Although this approach simplifies data acquisition compared to supervised methods, it struggles with reflective surfaces, as they violate the assumptions of Lambertian reflectance, leading to inaccurate training on such surfaces. To tackle this problem, we propose a novel training strategy for an SSMDE by leveraging triplet mining to pinpoint reflective regions at the pixel level, guided by the camera geometry between different viewpoints. The proposed reflection-aware triplet mining loss specifically penalizes the inappropriate photometric error minimization on the localized reflective regions while preserving depth accuracy in non-reflective areas. We also incorporate a reflection-aware knowledge distillation method that enables a student model to selectively learn the pixel-level knowledge from reflective and non-reflective regions. Evaluation results on multiple datasets demonstrate that our method effectively enhances depth quality on reflective surfaces and outperforms state-of-the-art SSMDE baselines. This approach significantly simplifies data acquisition compared to traditional supervised methods (Fu et al., 2018; Lee et al., 2019; Bhat et al., 2021), which often involve high costs for annotation. As such, many SSMDE studies (Godard et al., 2019; Zhou et al., 2017; Garg et al., 2016; Guizilini et al., 2020) have explored its viability as a mainstay for applications such as autonomous driving, highlighting its potential in outdoor environments. Despite its advantages, SSMDE approaches typically challenge in accurate depth estimation on non-Lambertian surfaces such as mirrors, transparent objects, and specular surfaces. This difficulty primarily arises from the assumption of Lambertian reflectance (Basri & Jacobs, 2003) embedded in most SSMDE methods.

This book provides insights into research in the field of artificial intelligence in combination with robotics technologies. The integration of artificial intelligence and robotic technologies is a highly topical area for researchers and developers from academia and industry around the globe, and it is likely that artificial intelligence will become the main approach for the next generation of robotics research. The tremendous number of artificial intelligence algorithms and big data solutions has significantly extended the range of potential applications for robotic technologies, and has also brought new challenges for the artificial intelligence community. Sharing recent advances in the field, the book features papers by young researchers presented at the 4th International Symposium on Artificial Intelligence and Robotics 2019 (ISAIR2019), held in Daegu, Korea, on August 20–24, 2019.

The COVID-19 pandemic has arguably been our era's greatest threat to humanity and the global economy.16 South Korea's first confirmed case was in January 2020, followed by an outbreak in the city of Daegu in mid-February. However, South Korea quickly and effectively contained the pandemic and became an exemplar for other countries.3 While many policies and initiatives contributed to South Korea's successful response to the coronavirus pandemic, digital technology was at the core of the endeavors.12 As part of its 3T strategy (test, trace, and treat) for coping with COVID-19, South Korea deployed a software system that traces the contacts of infected patients and disseminates the information in a matter of minutes.19 The COVID-19 Contact Tracing System CCTS) was first released in March 2020 to the Korea Centers for Disease Control and Prevention (KCDC)--a government agency responsible for advancing public health--and was then rolled out nationally in early April 2020. The system greatly contributed to reducing the number of daily new confirmed cases from 909 on February 29, 2020 to 7.42 on average between April 29 and May 5. According to a recent Columbia University study,13 both South Korea and the U.S. confirmed their first case of coronavirus on the same day. However, as of March 2021, South Korea's total confirmed cases are less than 10,000 and its proportional mortality rate is 50 times smaller than that of the U.S.7 The CCTS helped public healthcare officials to make informed decisions and helped keep the public aware of high-risk places where there had been exposure to coronavirus. Information provided by the CCTS enabled citizens to avoid hot spots and plan outdoor activities accordingly.

Tiny robots that can transport individual neurons and connect them to form active neural circuits could help us study brain disorders such as Alzheimer's disease. The robots, which were developed by Hongsoo Choi at the Daegu Gyeongbuk Institute of Science and Technology in South Korea and his colleagues, are 300 micrometres long and 95 micrometre wide. They are made from a polymer coated with nickel and titanium and their movement can be controlled with external magnetic fields.

After a passenger infected with the novel coronavirus boarded the Diamond Princess cruise ship in January, the virus quickly spread, eventually infecting at least 712 and killing seven. Critics labeled the ship quarantined in Yokohama a floating petri dish, and at least one Japanese expert attributed the explosion of cases to food trays passed out by infected crew. Could robots have made a difference? As countries around the world grapple with COVID-19, front line medical workers are deploying robots, telemedicine and other technologies to help contain the pandemic. China and Spain have used drones to monitor people during lockdown campaigns, while South Korea has deployed them to help disinfect areas in Daegu, an epidemic hotspot.

DGIST announced on Tuesday, July 16 that Senior Researcher Dae-gun Oh's team in the Collaborative Robots Research Center developed a radar system that can detect subminiature drones that are 3km away. This research is expected to make huge contributions to strengthening domestic industries and defense capabilities by securing a world-class radar sensing technology. As a result of discovering a North Korean drone in Paju in March 2014, South Korea's Ministry of National Defense has adopted a drone detection radar based on an overseas technology. Since last year, the ministry has devoted itself into building a combat system using drones and training specialized personnel by forming a drone unit to strengthen its defense capability. The necessity of enemy surveillance reconnaissance and the early detection of offensive drones has increased in Korea.

South Korean telecommunications carrier KT has developed a reader for its self-developed cellular vehicle-to-everything (C-V2X) technology, the company announced. KT said that the reader, when installed onto autonomous vehicles, will allow cars to detect passersby and traffic signals. The carrier also said the reader will allow cars to measure its distance from other vehicles to prevent potential collisions. The readers will be installed and tested in: Seoul, South Korea's capital; Pangyo, the country's equivalent to the US's Silicon Valley; and the southern city of Daegu. Once testing is complete, KT plans to commercialise the reader and put it to market.

In early 2015, a University of California, San Diego team successfully used micro-motor powered nanobots inside live mice -- without causing damage to their stomach linings, changing healthcare forever. In mid-2015, this concept was quickly advanced by mechanical engineers at Drexel University working in partnership with Daegu Gyeongbuk Institute of Science and Technology (DGIST) in South Korea. What they created were more efficient'micro-swimmers' capable of breaking through clogged arteries and leaving anticoagulant medication to prevent future blockage. Indeed, artificial intelligence (AI), from big data and machine learning to caretaker robots and medical nanobots, can help humans live longer. It's a primary reason scientists have predicted human lifespan to increase to 125 years by 2070.

The "Watson sensation" that hit Korea's medical industry seems to be waning. The domestic hospitals competed to introduce the artificial intelligence program of IBM last year, but no more joined the boom this year. Starting with Gachon University Gil Medical Center which introduced the AI system in December 2016, eight hospitals jumped on the Watson bandwagon last year --Pusan National University Hospital, Konyang University Hospital, Keimyung University Dongsan Medical Center, Daegu Catholic University Medical Center, Chosun University Hospital, Chonnam National University Hospital and VHS Medical Center. Some attribute the lack of followers to skepticism about the AI system, which had grown since the U.S. MD Anderson terminated its contract with IBM Watson last year. More point out, however, that it is the systemic limitations that cooled down the Watson fervor in Korea.

A research team at Korea's Daegu Gyeongbuk Institute of Science and Technology (DGIST) succeeded in analyzing big data up to 1,000 times faster than existing technology by using GPU-based'GMiner' technology. The finding of big data pattern analysis is expected to be utilized in various industries including the finance and IT sectors. An international team of researchers, led by Professor Min-Soo Kim from Department of Information and Communication Engineering developed'GMiner' technology that can analyze big data patterns at high speed. GMiner technology exhibits performance up to 1,000 times faster than the world's current best pattern mining technology. Pattern mining technology identifies all important patterns that appear repeatedly in the big data of various fields such as buying goods at mega-marts, banking transactions, network packets, and social networks.