Three Breakthroughs Driving the Rapid Rise of Computer Vision

Self‑Supervised Visual Foundation Models

Self‑supervised learning (SSL) learns from raw data without manual labels, enabling training of large‑parameter models. In vision, SSL yields Vision Foundation Models (VFMs) trained on massive unlabeled image collections. Two main SSL paradigms:

Generative SSL – e.g., Masked Auto‑Encoder (MAE, CVPR 2022) masks image patches and reconstructs them.

Contrastive SSL – e.g., SimCLR (ICML 2020) creates two augmented views per image and pulls together positive pairs while pushing apart negatives.

Meta’s DINO series illustrates vision SSL progression:

DINOv1 (ICCV 2021) introduces self‑distillation with a teacher‑student setup using global and local views.

DINOv2 (2023) scales training data to 0.14 B images and parameters to 1.1 B, enabling few‑shot fine‑tuning.

DINOv3 (2025) expands data to 1.7 B images and parameters to 6.7 B, achieving zero‑shot performance that surpasses supervised models on dense (segmentation, depth) and sparse (classification, detection) tasks.

Feed‑Forward 3D Reconstruction

Image‑based 3D reconstruction recovers 3D structure from 2D images. Traditional pipelines (Multi‑View Stereo, Structure‑from‑Motion, depth estimation) involve feature extraction, matching, pose estimation, triangulation, bundle adjustment, and dense reconstruction.

Feed‑forward approaches predict all 3D information in a single forward pass, removing iterative optimization.

VGGT (CVPR 2025) – a streamlined Transformer that ingests multi‑view images and simultaneously predicts camera parameters, point clouds, and depth maps.

VGGT‑Ω (CVPR 2026) – an improved version with more training data and architectural refinements.

MapAnything (2026) – a universal feed‑forward metric 3D reconstruction model using a discrete scene representation; accepts images plus optional sensor data and directly predicts real‑world scale factors.

Implicit and explicit novel‑view synthesis methods:

NeRF (ECCV 2020) represents a scene with an MLP that outputs density and color for any spatial coordinate; requires per‑scene training.

3D Gaussian Splatting (SIGGRAPH 2023) uses an explicit set of anisotropic Gaussians with differentiable rasterization, achieving real‑time rendering.

Unified Multimodal Models

Vision‑Language Models (VLMs) are categorized as:

Multimodal Large Language Models (MLLMs) – e.g., Qwen‑VL, LLaVA, which treat visual input as a perception token for a large language model, enabling zero‑shot visual understanding and interactive QA.

Visual‑backbone VLMs – e.g., SAM series, which keep the vision model as the primary component and use text prompts to guide segmentation.

Two newer families address limitations of early MLLMs:

Native Multimodal Models (NMM) – train vision and text jointly from scratch (e.g., Qwen 3.5), eliminating separate pre‑trained modules.

Unified Multimodal Models (UMM) – support both multimodal understanding and generation (image, video synthesis) via external expert, modular, or end‑to‑end designs.

Meta’s SAM3 (2025) extends the Segment‑Anything paradigm with promptable concept segmentation, supporting image/video inputs, text prompts, points, boxes, and reference images; scales to over 4 M fine‑grained concepts with 8.5 M parameters.

References

[1] He K. et al., “Masked Autoencoders are Scalable Vision Learners,” CVPR 2022.

[2] Chen T. et al., “A Simple Framework for Contrastive Learning of Visual Representations,” ICML 2020.

[3] Caron M. et al., “Emerging Properties in Self‑Supervised Vision Transformers,” ICCV 2021.

[4] Oquab M. et al., “Dinov2: Learning Robust Visual Features without Supervision,” arXiv 2023.

[5] Siméoni O. et al., “Dinov3,” arXiv 2025.

[6] Morelli L. et al., “COLMAP‑SLAM: A Framework for Visual Odometry,” 2023.

[7] Rosinol A. et al., “Kimera: An Open‑Source Library for Real‑Time Metric‑Semantic Localization and Mapping,” ICRA 2020.

[8] Lin H. et al., “Depth Anything 3: Recovering the Visual Space from Any Views,” arXiv 2025.

[9] Mildenhall B. et al., “NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis,” CACM 2021.

[10] Kerbl B. et al., “3D Gaussian Splatting for Real‑Time Radiance Field Rendering,” ACM TOG 2023.

[12] Wang J. et al., “VGGT: Visual Geometry Grounded Transformer,” CVPR 2025.

[13] Keetha N. et al., “MapAnything: Universal Feed‑Forward Metric 3D Reconstruction,” 2026.

[15] Carion N. et al., “SAM 3: Segment Anything with Concepts,” arXiv 2025.

Code example

[2]Chen T, Kornblith S, Norouzi M, et al. A simple framework for contrastive learning of visual representations[C]//International conference on machine learning. PmLR, 2020: 1597-1607.
[3]Caron M, Touvron H, Misra I, et al. Emerging properties in self-supervised vision transformers[C]//Proceedings of the IEEE/CVF international conference on computer vision. 2021: 9650-9660.
[4]Oquab M, Darcet T, Moutakanni T, et al. Dinov2: Learning robust visual features without supervision[J]. arXiv preprint arXiv:2304.07193, 2023.
[5]Siméoni O, Vo H V, Seitzer M, et al. Dinov3[J]. arXiv preprint arXiv:2508.10104, 2025.