💷 SOTA Zero-Shot Stereo Matching💷

👉Fast-FoundationStereo by #Nvidia is a novel family of architectures that achieve, for the first time, strong zero-shot generalization at real-time frame rate via divide-&-conquer acceleration. Code & Data announced💙

👉Review https://t.ly/XD6pO
👉Paper https://lnkd.in/d9_YKW2A
👉Project https://lnkd.in/dKDxm7EX
👉Repo https://lnkd.in/dR4-PdsW

👀DriverGaze360: Driver SOTA👀

👉DriverGaze360 is a large-scale 360◦ field of view driver attention dataset, containing ∼1M gaze-labeled frames. Code & Dataset announced💙

👉Review https://t.ly/ZcoUw
👉Paper arxiv.org/pdf/2512.14266
👉Project av.dfki.de/drivergaze360/
👉Repo github.com/dfki-av/drivergaze360
👉Data av.dfki.de/drivergaze360/dataset

🫠FlexAvatar: 3D Heads🫠

👉TUM introduces FlexAvatar, a novel method for creating HQ and complete 3D head avatars from a single image. Code announced💙

👉Review https://t.ly/Rkdtd
👉Paper arxiv.org/pdf/2512.15599
👉Project tobias-kirschstein.github.io/flexavatar/
👉Repo TBA

🏜️ Depth Any Panoramas 🏜️

👉DAP is the new SOTA foundation model for panoramic depth estimation with a large scale dataset. Data & Repo under MIT💙

👉Review https://t.ly/LaUmd
👉Paper arxiv.org/pdf/2512.16913
👉Project https://lnkd.in/dvqNV9jx
👉Repo https://lnkd.in/dmNzhb-7
👉Demo https://lnkd.in/dDwjMF3u

🎯Generative Refocusing is out🎯

👉Generative Refocusing is a two-step process that uses DeblurNet to recover all-in-focus images from various inputs and BokehNet for creating controllable bokeh (in semi-supervised mode). Repo under Apache2.0💙

👉Review https://t.ly/8t7PA
👉Paper arxiv.org/pdf/2512.16923
👉Project generative-refocusing.github.io/
👉Repo github.com/rayray9999/Genfocus
👉Demo huggingface.co/spaces/nycu-cplab/Genfocus-Demo

⭐TOP 5 Papers you loved in 2025⭐

👉 In 2025 novel architectures have redefined efficiency and accuracy, and almost every day brought a new SOTA in image understanding, tracking, and GenAI. It’s been an inspiring ride, and 2026 it will be even wilder. This community (LinkedIn + Telegram) is now around 80,000+ people.

𝐏𝐚𝐩𝐞𝐫𝐬 (𝐛𝐲 𝐲𝐨𝐮𝐫 𝐩𝐫𝐞𝐟𝐞𝐫𝐞𝐧𝐜𝐞):
⭐3D LLM https://t.ly/ejr1s
⭐DynOMo https://t.ly/t5pCf
⭐Track Transf. https://t.ly/NPyW4
⭐YOLOv12 https://t.ly/jj1oR
⭐G-Surface Tracking https://t.ly/udpMq

Thank you all💙

🦙 Depth as Neural Implicit 🦙

👉InfiniDepth represents depth as neural implicit fields, "infinite" (i.e.16K) resolution and geometrical details. Repo under Apache 2.0💙

👉Review https://t.ly/4we5t
👉Paper https://lnkd.in/dpiHQExj
👉Project https://lnkd.in/dy3JxKye
👉Repo https://lnkd.in/dAXbnK5z

🔥 Back from Holidays mood 🔥

🌍Label Any Object in 3D 🌍

👉LabelAny3D: novel analysis-by-synthesis framework that reconstructs holistic 3D scenes from 2D to efficiently produce HQ 3D BBs annotations. Repo under CC-BY-4.0 license💙

👉Review https://t.ly/bO93j
👉Paper https://lnkd.in/dYb97zWG
👉Project https://lnkd.in/dJ9UKERb
👉Repo https://lnkd.in/d9SxtmiA

🔥 New #AI Startups in 2026? 🔥

In 2026, which area would you focus on?
🤖Agents → workflows, copilots, etc.
🏭Vertical AI → Pharma, Automotive, Energy ...
🧠Infrastructure → MLOps, Security, Cost Control ...
🎨AI for Creators/Media → Video, avatars, contents ...

Please, help me understanding what's next with this poll on LinkedIn :)

https://www.linkedin.com/posts/visionarynet_ai-ai-deeplearning-activity-7415377341779996672-sQO1

LUV U \m/

🔥Orient Anything V2 is out🔥

👉Orient Anything V2 is a foundation model for unified understanding of object 3D orientation and rotation from single or paired images. Repo under CC-BY-4.0💙

👉Review https://t.ly/Ht7Xd
👉Paper arxiv.org/pdf/2601.05573
👉Project orient-anythingv2.github.io/
👉Repo github.com/SpatialVision/Orient-Anything-V2

🫛Active Object Reconstruction🫛

👉ObjSplat (Beijing) autonomously plans viewpoints and progressively reconstructs an unknown object into a Hi-Fi Gaussian model and water-tight mesh, enabling direct use in physics simulations. Tough paper and repo announced💙

👉Review https://t.ly/au6HE
👉Paper arxiv.org/pdf/2601.06997
👉Project li-yuetao.github.io/ObjSplat-page/
👉Repo https://github.com/Li-Yuetao/ObjSplat

In 2026, who should we keep an eye on?

Vote: https://www.linkedin.com/posts/visionarynet_ai-deeplearning-aiwithpapers-activity-7416886610795077632-qQeP/

👉Games Workshop (Warhammer) is banning the use of AI in creative and design processes to protect IP and human creativity. A decision that goes against the current hype of widespread AI adoption.

And what about your organization? I need your help👇

Vote: https://www.linkedin.com/posts/visionarynet_ai-activity-7417106327019196417-TpGL

💚Segment Anything Geometry💚

👉3AM (NYCU + #Nvidia) offers cross-view correspondence even under large viewpoint changes, cluttered scenes, and variations in capture conditions, enabling robust object tracking from both videos & casual multi-view images. Repo (coming) & Demo available💙

👉Review https://t.ly/olZwE
👉Paper https://arxiv.org/pdf/2601.08831
👉Project https://jayisaking.github.io/3AM-Page/
👉Repo https://github.com/jayisaking
👉Demo https://huggingface.co/spaces/nycu-cplab/3AM

🎇 Multi-target SAM3 🎇

👉SAM3-DMS is a novel training-free decoupled strategy that utilizes fine-grained memory selection on individual objects. Robust identity preservation and tracking stability. Repo under SAM License💙

👉Review https://t.ly/jJOAr
👉Paper https://arxiv.org/pdf/2601.09699
👉Repo https://github.com/FudanCVL/SAM3-DMS

🍿100M Video Action Dataset🍿

👉Action100M by META is a large-scale dataset w/ 1.2M instructional videos (14.6 years of duration), yielding O(100M) temporally localized segments with open-vocabulary action supervision and rich captions. Repo under FAIR NC Research License💙

👉Review https://t.ly/w5KXe
👉Paper arxiv.org/pdf/2601.10592
👉Repo github.com/facebookresearch/Action100M

💜Interactive Humanoid Generation💜

👉FlowAct-R1 by ByteDance is a novel framework that enables lifelike, responsive, and high-fidelity humanoid video generation for seamless real-time interaction. No code but impressive results (see video with audio) 💙

👉Review https://t.ly/aQhol
👉Paper arxiv.org/pdf/2601.10103
👉Project grisoon.github.io/FlowAct-R1/

💢3D Human Gen-Seg💢

👉CoMoVi takes an input image with a text description and generates 3D human motion & video sequence synchronously within a single diffusion denoising loop. Repo & Dataset releasing💙

👉Review https://t.ly/khSkm
👉Paper arxiv.org/pdf/2601.10632
👉Project igl-hkust.github.io/CoMoVi/
👉Repo github.com/IGL-HKUST/CoMoVi
👉Data huggingface.co/datasets/AfterJourney/CoMoVi-Dataset

👹SOTA Part-level Generator👹

👉A novel a text-to-motion model that learns to compose complex motions through hierarchical conditioning on part-, action- & sequence-level text, enabling fine-grained control over body parts & timing. Code, models & Dataset to be released💙

👉Review https://t.ly/leB_R
👉Paper arxiv.org/pdf/2601.10909
👉Project coral79.github.io/frankenmotion/
👉Repo github.com/Coral79/FrankenMotion-Code