논문리뷰 - Choonsikmom Archive

[논문리뷰] DINOv3: Self-supervised Learning for Vision at Scale (Meta AI, 2025)

August 19, 2026

7B ViT와 Gram anchoring으로 dense feature 붕괴를 잡고 frozen backbone만으로 종합 SOTA를 달성한 SSL foundation model

DINOv3 self-supervised learning ViT Gram anchoring knowledge distillation vision foundation model

[논문리뷰] DINOv2: Learning Robust Visual Features without Supervision (TMLR 2024)

August 18, 2026

curated LVD-142M 데이터와 개선된 iBOT 기반 recipe로 ViT-g(1.1B)를 학습하고 distillation로 전파해 OpenCLIP를 넘어선 범용 visual feature

DINOv2 self-supervised learning ViT data curation knowledge distillation

[논문리뷰] DINO: Emerging Properties in Self-Supervised Vision Transformers (ICCV 2021)

August 17, 2026

label 없는 self-distillation로 ViT를 학습해 attention map에 segmentation이 창발하고 k-NN 분류 78.3%에 도달하는 DINO

DINO self-supervised learning ViT knowledge distillation representation learning

[논문리뷰] SAM 3: Segment Anything with Concepts (ICLR 2026)

August 15, 2026

noun phrase와 exemplar prompt로 concept의 모든 instance를 검출·분할·추적하는 promptable concept segmentation 모델 SAM 3 논문 리뷰

open-vocabulary segmentation promptable segmentation video segmentation object tracking foundation model

[논문리뷰] Khronos: A Unified Approach for Spatio-Temporal Metric-Semantic SLAM in Dynamic Environments (RSS 2024)

August 13, 2026

단기 dynamics와 장기 변화를 object fragment factorization으로 통합해 실시간 4D spatio-temporal map을 만드는 Khronos 논문 리뷰

SLAM dynamic scene understanding change detection factor graph metric-semantic mapping

[논문리뷰] GenZ-ICP: Generalizable and Degeneracy-Robust LiDAR Odometry Using an Adaptive Weighting (RA-L 2025)

August 13, 2026

point-to-plane과 point-to-point residual을 평면성 비율 기반 adaptive weight로 결합해 복도 degeneracy에 강건한 LiDAR odometry 논문 리뷰

LiDAR odometry ICP point cloud registration degeneracy SLAM

[논문리뷰] FoundationStereo: Zero-Shot Stereo Matching (CVPR 2025)

August 12, 2026

1M 합성 데이터셋과 monocular prior 적응, disparity transformer로 zero-shot stereo matching을 실현한 논문 리뷰

stereo matching depth estimation zero-shot generalization cost volume synthetic data

[논문리뷰] 3DGS-CD: 3D Gaussian Splatting-based Change Detection for Physical Object Rearrangement (RA-L 2025)

August 7, 2026

3DGS novel view 렌더링과 EfficientSAM zero-shot 분할로 재배치된 물체의 3D mask와 6D pose 변화를 추정하는 change detection 논문 리뷰

3D change detection 3D Gaussian Splatting EfficientSAM object rearrangement pose estimation

[논문리뷰] Living Scenes: Multi-object Relocalization and Reconstruction in Changing 3D Environments (CVPR 2024)

August 5, 2026

SE(3)-equivariant 표현 하나로 시간차 scan의 instance matching·registration·reconstruction을 함께 푸는 MoRE² 논문 리뷰

3D scene understanding point cloud registration SE(3) equivariance neural implicit representation instance matching

[논문리뷰] FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects (CVPR 2024)

August 5, 2026

CAD model 유무와 무관하게 novel object의 6D pose estimation과 tracking을 하나의 framework로 통합한 FoundationPose 논문 리뷰

6D pose estimation object pose tracking neural implicit representation synthetic data generation contrastive learning

[논문리뷰] π³: Permutation-Equivariant Visual Geometry Learning (ICLR 2026)

August 3, 2026

기준 view를 없애고 permutation-equivariant 구조로 affine-invariant pose와 scale-invariant point map을 예측하는 π³ 논문 리뷰

visual geometry 3D reconstruction camera pose estimation permutation equivariance depth estimation

[논문리뷰] MapAnything: Universal Feed-Forward Metric 3D Reconstruction (3DV 2026)

August 3, 2026

이미지와 선택적 geometric 입력에서 factored representation으로 metric 3D를 직접 회귀하는 통합 feed-forward 모델 MapAnything 논문 리뷰

MapAnything 3D reconstruction metric depth estimation multi-view geometry structure from motion

[논문리뷰] VGGT: Visual Geometry Grounded Transformer (CVPR 2025)

July 31, 2026

camera·depth map·point map·3D point track을 단일 feed-forward transformer의 forward 한 번으로 예측하는 3D vision 통합 모델 VGGT 논문 리뷰

VGGT 3D reconstruction camera pose estimation multi-view geometry point tracking

[논문리뷰] GoTrack: Generic 6DoF Object Pose Refinement and Tracking (CVPRW 2025)

July 29, 2026

optical flow로 model-to-frame과 frame-to-frame registration을 결합해 unseen object의 6DoF pose를 정제·추적하는 GoTrack 논문 리뷰

6DoF pose estimation object tracking pose refinement optical flow DINOv2

[논문리뷰] CoTracker3: Simpler and Better Point Tracking by Pseudo-Labelling Real Videos (ICCV 2025)

July 29, 2026

기존 tracker들을 teacher로 삼아 real video에 pseudo-label을 달고 1,000배 적은 데이터로 SOTA를 넘는 단순화된 point tracker CoTracker3 논문 리뷰

CoTracker point tracking semi-supervised learning pseudo-labeling knowledge distillation

[논문리뷰] CoTracker: It is Better to Track Together (ECCV 2024)

July 29, 2026

track 간 attention과 support point로 여러 점을 jointly 추적해 occlusion에 강한 transformer 기반 point tracker CoTracker 논문 리뷰

CoTracker point tracking optical flow motion estimation transformer

[논문리뷰] Depth Anything 3: Recovering the Visual Space from Any Views (2025)

July 26, 2026

depth와 ray map을 최소 prediction target으로 삼아 plain transformer 하나로 any-view geometry를 통일한 Depth Anything 3 논문 리뷰

Depth Anything multi-view geometry camera pose estimation monocular depth estimation 3D Gaussian Splatting knowledge distillation

[논문리뷰] Depth Anything V2 (NeurIPS 2024)

July 26, 2026

실측 label을 전부 synthetic 이미지로 바꾸고 62M pseudo-labeled 실사 이미지로 증류해 정밀도와 robustness를 함께 잡은 Depth Anything V2 논문 리뷰

Depth Anything monocular depth estimation synthetic data knowledge distillation pseudo label zero-shot

[논문리뷰] Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data (CVPR 2024)

July 18, 2026

62M unlabeled 이미지 self-training과 DINOv2 feature alignment로 만든 zero-shot MDE foundation model, Depth Anything 논문 리뷰

Depth Anything monocular depth estimation self-training semi-supervised learning DINOv2 zero-shot

[논문리뷰] Grounding Image Matching in 3D with MASt3R (ECCV 2024)

July 14, 2026

matching을 3D 문제로 정식화해 DUSt3R에 local feature head와 InfoNCE loss를 더하고 Map-free localization에서 30%p 개선을 이룬 MASt3R 논문 리뷰

MASt3R image matching 3D Reconstruction visual localization DUSt3R

[논문리뷰] DUSt3R: Geometric 3D Vision Made Easy (CVPR 2024)

July 14, 2026

카메라 calibration도 pose도 없이 이미지 쌍에서 pointmap을 직접 회귀해, depth·대응·pose·3D 복원까지 기하 3D 비전 과제 전체를 하나의 모델로 푸는 DUSt3R 논문 리뷰

DUSt3R 3D Reconstruction pointmap Transformer MVS

[논문리뷰] FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (NeurIPS 2022)

July 13, 2026

attention의 병목을 FLOPs가 아닌 HBM 접근으로 규정하고 tiling·recomputation으로 O(N) 메모리의 exact attention을 구현한 FlashAttention 논문 리뷰

FlashAttention attention Transformer GPU tiling

[논문리뷰] SAM-6D: Segment Anything Model Meets Zero-Shot 6D Object Pose Estimation (CVPR 2024)

July 12, 2026

SAM 기반 ISM과 background token point matching의 PEM으로 zero-shot 6D pose를 구하는 SAM-6D 논문 리뷰

SAM-6D 6D pose estimation zero-shot SAM point matching

[논문리뷰] PPO: Proximal Policy Optimization Algorithms (2017)

July 5, 2026

clipped surrogate objective로 TRPO의 안정성을 1차 최적화만으로 달성한 PPO 논문 리뷰

PPO reinforcement learning policy gradient TRPO