Research 2024.12.03

Joint Appearance and Motion Model with Temporal Transformer for Multiple Object Tracking

(English Only) The problem of multi-object tracking (MOT) in the real world poses several challenging tasks, such as similar appearance, occlusion, and extreme articulation motion.

Blog 2024.09.02

Active Learning을 통한 지속적인 모델 성능 개선

42dot에서는 실시간으로 작동하는 고성능 3차원 인지 모델을 개발하기 위해 지속적으로 데이터를 수집하고, 모델의 성능을 추적하여 자율주행 기술을 발전시키고 있습니다.

Research 2024.05.20

Towards Understanding the Relationship between In-context Learning and Compositional Generalization

본 논문에서는 in-context learning을 모델에 강제하는 것이 구성적 일반화를 촉진하는 귀납적 편향(inductive bias)을 제공할 수 있다는 가설을 제시합니다.

Research 2024.04.24

Learning Contextualized Representation On Discrete Space Via Hierarchical Product Quantization

(English Only) According to the principle of compositional generalization, the meaning of a complex expression can be understood as a function of the meaning of its parts and of how they are combined.

Research 2024.04.04

Boosting Unknown-number Speaker Separation With Transformer Decoder-based Attractor

(English Only) We propose a novel speech separation model designed to separate mixtures with an unknown number of speakers.

Research 2024.04.04

Voxtlm: Unified Decoder-only Models for Consolidating Speech Recognition/Synthesis and Speech/Text Continuation Tasks

(English Only) We propose a decoder-only language model, \textit{VoxtLM}, that can perform four tasks: speech recognition, speech synthesis, text generation, and speech continuation.

Blog 2024.03.29

42dot LLM 1.3B

42dot에서는 지난 가을, 자체 개발한 초거대 언어 모델인 42dot LLM을 공개한 바 있습니다. 42dot LLM은 국내 최초의 한영 통합 언어 모델로서 직접 구축한 데이터와 자체 학습 인프라를 활용해, 비슷한 규모의 다른 언어 모델 대비 월등한 성능을 달성하며 좋은 품질을 보여줬습니다.

Blog 2024.03.28

42dot at CES 2024: Software-Defined Vehicle Technology

CES 2024에서 현대자동차그룹이 발표한 ‘software-defined everything’ 전략에 맞춰 그룹의 글로벌 소프트웨어 센터인 42dot이 공개한 새로운 SDV 전기・전자 아키텍처와 핵심 기술들을 소개합니다.

Research 2023.10.31

TF-GridNet: Integrating Full- and Sub-Band Modeling for Speech Separation

(English Only) We propose TF-GridNet for speech separation. The model is a novel deep neural network (DNN) integrating full- and sub-band modeling in the time-frequency (T-F) domain.

Research 2023.10.30

That's What Said: Fully-Controllable Talking Face Generation

(English Only) The goal of this paper is to synthesise talking faces with controllable facial motions. To achieve this goal, we propose two key ideas. The first is to establish a canonical space where every face has the same motion patterns but different identities.

Research 2023.10.24

SpeedFormer: Learning Speed Profiles with Upper and Lower Boundary Constraints Based on Transformer

(English Only) This paper presents a new method for generating speed profiles for autonomous vehicles using a Transformer-based network that predicts the coefficients of quintic polynomials.

Research 2023.10.02

SlaBins: Fisheye Depth Estimation using Slanted Bins on Road Environments

(English Only) Although 3D perception for autonomous vehicles has focused on frontal-view information, more than half of fatal accidents occur due to side impacts in practice (e.g., T-bone crash).

Research 2023.10.02

Luminance-aware Color Transform for Multiple Exposure Correction

(English Only) Images captured with irregular exposures inevitably present unsatisfactory visual effects, such as distorted hue and color tone.

Open Dataset 2023.09.22

42dot LLM-SFT

42dot LLM-SFT는 42dot에서 개발한 LLM(거대 언어 모델)으로 자연어 instruction을 따르도록 학습되었습니다. 42dot LLM-SFT는 42dot LLM의 일부이며 supervised fine-tuning (SFT)을 통해 42dot LLM-PLM에서 파생되었습니다. 이 저장소에는 1.3B parameter 버전이 포함되어 있습니다.

Open Dataset 2023.09.13

42dot LLM-PLM

42dot-PLM은 42dot에서 개발한 PLM (pre-trained language model)으로 42dot LLM(거대 언어 모델)의 일부입니다. 42dot LLM-PLM은 한국어 및 영어 text corpus를 사용하여 사전 학습되었으며 여러 자연어 처리 태스크 작업의 기초 언어 모델로 사용할 수 있습니다. 이 저장소에는 모델의 1.3B-parameter 버전이 포함되어 있습니다.

Blog 2023.09.12

영지식 증명과 블록체인 그리고 SDV, 모빌리티

42dot에서는 모빌리티 사용자에게 새로운 유형의 모틸리티 경험을 제공하기 위한 블록체인 플랫폼을 개발하고 있습니다. 영지식 증명은 42dot이 생각하고 있는 블록체인의 중요한 기술적인 기반 중 하나 입니다.

Research 2023.08.24

Factspeech: Speaking a Foreign Language Pronunciation Using Only Your Native Characters

(English Only) Recent text-to-speech models have been requested to synthesize natural speech from language-mixed sentences because they are commonly used in real-world applications. However, most models do not consider transliterated words as input.

Open Dataset 2023.07.04

MCMOT: multi-camera multi-object tracking

자율주행 시스템은 여러 대의 카메라를 활용해 주변 환경을 인식합니다. 따라서 카메라 시점을 가로질러 움직이는 객체를 인식하려면 동일한 트랙 ID를 유지하는 것이 중요합니다. 우리는 3개의 전면 카메라로 캡처한 객체에 고유한 트랙 ID를 할당하는 주석이 달린 데이터 세트를 제공합니다.

Blog 2023.06.30

Team 42dot Wins 2nd Place in the Autonomous Driving Challenge at CVPR 2023

(English Only) 42dot Inc. has presented the solution referred to as MiLO which won the 2nd place (honorable runner-up) in the fiercely contested 3D Occupancy Prediction Challenge for autonomous driving at the Computer Vision and Pattern Recognition Conference (CVPR) 2023; in Vancouver, Canada.

Research 2023.06.19

MiLO: Multi-task Learning with Localization Ambiguity Suppression for Occupancy Prediction

(English Only) We present Multi-task Learning with Localization Ambiguity Suppression for Occupancy Prediction (MiLO) as our solution for camera-based 3D Occupancy Prediction Challenge at CVPR 2023.

Research 2023.06.19

BAAM: Monocular 3D pose and shape reconstruction with bi-contextual attention module and attention-guided modeling

(English Only) A novel monocular 3D pose and shape reconstruction algorithm, based on bi-contextual attention and attention-guided modeling (BAAM), is proposed in this work.

Research 2023.06.19

RUFI: Reducing Uncertainty in behavior prediction with Future Information

(English Only) Autonomous driving has shown significant progress in recent years, but accurately predicting the movements of surrounding traffic agents remains a challenge for ensuring safety. Previous studies have focused on behavior prediction using large-scale data with diverse information like lane and agent information. 

Research 2023.06.06

Masked Token Similarity Transfer for Compressing Transformer-Based ASR Models

(English Only) Recent self-supervised automatic speech recognition (ASR) models based on transformers are showing best performance, but their footprint is too large to be trained on low-resource environments or deployed to edge devices. Knowledge distillation (KD) can be employed to reduce the model size.

Research 2023.06.05

Metric Learning for User-defined Keyword Spotting

(English Only)The goal of this work is to detect new spoken terms defined by users. While most previous works address Keyword Spotting (KWS) as a closed-set classification problem, this limits their transferability to unseen terms.

Research 2023.06.05

TF-GridNet: Making Time-Frequency Domain Models Great Again for Monaural Speaker Separation

(English Only) We propose TF-GridNet, a novel multi-path deep neural network (DNN) operating in the time-frequency (T-F) domain, for monaural talker-independent speaker separation in anechoic conditions.

Research 2023.06.05

Joint unsupervised and supervised learning for context-aware language identification

(English Only) Language identification (LID) recognizes the language of a spoken utterance automatically. According to recent studies, LID models trained with an automatic speech recognition (ASR) task perform better than those trained with a LID task only.

Research 2023.06.05

Neural Speech Enhancement with Very Low Algorithmic Latency and Complexity via Integrated Full- and Sub-Band Modeling

(English Only) We propose FSB-LSTM, a novel long short-term memory (LSTM) based architecture that integrates full- and sub-band (FSB) modeling, for single- and multi-channel speech enhancement in the short-time Fourier transform (STFT) domain.

Research 2023.06.05

CrossSpeech: Speaker-independent Acoustic Representation for Cross-lingual Speech Synthesis

(English Only) While recent text-to-speech (TTS) systems have made remarkable strides toward human-level quality, the performance of cross-lingual TTS lags behind that of intra-lingual TTS. This gap is mainly rooted from the speaker-language entanglement problem in cross-lingual TTS.

Blog 2023.04.19

Joint Unsupervised and Supervised Learning for Context-aware Language Identification

2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2023)에서 발표한 “Joint unsupervised and supervised learning for context-aware language identification” 논문을 소개합니다.

Research 2023.01.09

ASBERT: ASR-Specific Self-Supervised Learning with Self-Training

(English Only) Pre-training of self-supervised learning (SSL) generally shows a good performance on various speech processing tasks. However, this pre-training scheme may lead to a sub-optimal solution for fine-tuning a specific task, such as automatic speech recognition (ASR).

Blog 2023.01.05

AWS IoT Core Resource Deployment via CDK

AWS Cloud Development Kit(이하 AWS CDK)는 익숙한 프로그래밍 언어를 사용하여 클라우드 애플리케이션 리소스를 정의할 수 있는 오픈 소스 소프트웨어 개발 프레임워크입니다. 이러한 코드를 통해 인프라를 관리하는 방식을 Infrastructure as Code, 줄여서 IaC라고 부릅니다.

Blog 2022.12.16

ML Data Platform for Continuous Learning

42dot에서는 자율주행기술 개발을 위해 머신러닝 기술을 적극적으로 활용하고 있습니다. 머신러닝 개발의 경우 고도의 알고리즘, 대량의 데이터 그리고 복잡한 컴퓨팅 연산이 필요하고 이를 수행하기 위해서는 software 및 hardware의 효율적인 지원이 필요합니다. 이를 위에 42dot에서는 다양한 기술을 이용하여 machine learning과 data platform을 개발하여 운영하고 있습니다.

Research 2022.11.28

Self-supervised surround-view depth estimation with volumetric feature fusion

(English Only) We present a self-supervised depth estimation approach using a unified volumetric feature fusion for surround-view images. Given a set of surround-view images, our method constructs a volumetric feature map by extracting image feature maps from surround-view images and fuse the feature maps into a shared, unified 3D voxel space.

Blog 2022.11.25

속도와 보안이 강화된 OTA 업데이트

42dot에서는 모든 것들이 스스로 움직이게 하기 위해서 자율 주행 제어 장치를 개발하고 있습니다. 이 장치는 다양한 센서(카메라, 레이더 등)를 활용해서 주변 상황을 인지/판단하고 지도/측위 기술 등을 조합해서 주행 경로에 따라 장치 스스로 조향 각과 가감속과 같은 다양한 이동 과정의 제어를 수행합니다.

Research 2022.11.07

An Empirical Study of Training Mixture Generation Strategies on Speech Separation: Dynamic Mixing and Augmentation

(English Only) Deep learning has dramatically advanced speech separation (SS) in the past decade. Although advances in model architectures play an essential role in improving the separation performance, an efficient training strategy is also important.

Research 2022.10.24

Character decomposition to resolve class imbalance problem in Hangul OCR

(English Only) We present a novel approach to OCR(Optical Character Recognition) of Korean character, Hangul. As a phonogram, Hangul can represent 11,172 different characters with only 52 graphemes, by describing each character with a combination of the graphemes.

Blog 2022.10.14

Foros : 자동차에 합의 알고리즘을?

미래 기술이라 생각했던 자율주행차를 정식 교통수단으로 이용할 수 있는 시대가 되었습니다. 이에 따라 42dot 은 자율주행 안정성을 높이기 위한 연구 개발을 진행하고 있으며, 그중 ‘합의 알고리즘 기반 애플리케이션 다중화 기술’에 대해서 이야기하려고 합니다.

Blog 2022.09.20

42dot MCMOT(Multi-Camera Multi-Object Tracking) 챌린지

42dot이 자율주행 연구 개발을 위한 생태계 조성의 일환으로 공개한 자율주행 데이터 ‘42dot Open Dataset’을 토대로 진행한 MCMOT 챌린지 결과와 그 내용을 공개합니다.

Blog 2022.09.16

42dot이 그리는 미래 모빌리티 세상

자율주행이 가져올 우리 삶의 변화, 어떤 모습일지 상상해 본 적 있나요? 42dot이 미션인 ‘autonomous and frictionless'의 가치를 담아 모빌리티의 미래 모습을 영상으로 만들었습니다. 42dot이 기대하는 모빌리티의 미래, 함께 살펴보겠습니다.

Research 2022.06.20

Eigenlanes: Data-driven lane descriptors for structurally diverse lanes

(English Only) A novel algorithm to detect road lanes in the eigenlane space is proposed in this paper. First, we introduce the notion of eigenlanes, which are data-driven descriptors for structurally diverse lanes, including curved, as well as straight, lanes. 

Open Dataset 2022.06.02

42dot releases Open Dataset, SDLane

SDLane dataset은 자율주행을 위한 새로운 차선 dataset입니다. Dataset은 고속도로와 도심의 복잡한 시나리오를 담은 1920 X 1208 픽셀의 고해상도 이미지를 제공합니다. SD Lane dataset는 정확한 실측 라벨이 포함된 39K 훈련 이미지와 4K 테스트 이미지로 구성됩니다.

Research 2021.07.19

Harmonious semantic line detection via maximal weight clique selection

(English Only) A novel algorithm to detect an optimal set of semantic lines is proposed in this work. We develop two networks: selection network (S-Net) and harmonization network (H-Net). First, S-Net computes the probabilities and offsets of line candidates.

Research 2019.11.25

Anchor Loss: Modulating Loss Scale based on Prediction Difficulty

(English Only) We propose a novel loss function that dynamically re-scales the cross entropy based on prediction difficulty regarding a sample. Deep neural network architectures in image classification tasks struggle to disambiguate visually similar objects.

Research 2019.10.29

Instance-level future motion estimation in a single image based on ordinal regression

(English Only) A novel algorithm to estimate instance-level future motion in a single image is proposed in this paper. We first represent the future motion of an instance with its direction, speed, and action classes. Then, we develop a deep neural network that exploits different levels of semantic information to perform the future motion estimation.

Research 2019.10.28

Drop to Adapt: Learning Discriminative Features for Unsupervised Domain Adaptation

(English Only) Domain adversarial methods render suboptimal performances since they attempt to match the distributions among the domains without considering the task at hand. We propose Drop to Adapt (DTA), which leverages adversarial dropout to learn strongly discriminative features by enforcing the cluster assumption.